{"slug": "qwen-3-8", "name": "Qwen 3.8", "category": "ai", "category_name": "AI models and assistants", "description": "Alibaba's open-weight language model family with 27B and Flash Next variants.", "recommenders": 64, "commenters": 107, "warners": 8, "score": 60.534, "rising": 64.0, "first": "2026-07-20", "latest": "2026-10-27", "why": "HN commenters run it locally on consumer GPUs and Macs as a daily driver, praising its coding ability for its size.", "why_is_quote": false, "has_verdict": true, "sentiment": {"recommend": 67, "neutral": 36, "compare": 4, "warn": 9}, "spark": [0, 0, 0, 0, 0, 0, 0, 0, 0, 25, 20, 22], "listed": true, "loved": ["Runs locally", "Coding ability"], "caveat": "Overthinking", "short_quote": "out of all the models i've used locally qwen3.8:27b blows everything out of the water.", "icon": true, "domain": "qwen.ai", "github_repo": null, "official_url": "https://qwen.ai", "pricing": "free", "aliases": ["Qwen 3.8 27B", "Qwen 3.8", "Qwen3.8 27B", "Qwen3.8-27B", "Qwen 3.8 Flash Next", "Qwen3.8-Flash-Next", "Qwen3.8", "Qwen Flash Next", "Qwen 3.8 Max", "Qwen3.8-Next-Flash", "Qwen-3.8-Flash-Next", "Qwen3.8 27b", "Qwen3.8 Flash Next", "Qwen3.8-flash", "Qwen3.8 Max", "Qwen 3.8 Flash", "Qwen3.8-Max", "Qwen 3.8 Max Preview", "Qwen3.8-Max-Preview", "Qwen 3.8 27b", "Qwen-3.8-27b", "Qwen-Flash-Next", "Qwen3.8-27b"], "verdict": [{"text": "HN commenters run it locally on consumer GPUs and Macs as a daily driver", "cites": [{"id": 803865472, "item_id": 50241592, "url": "https://news.ycombinator.com/item?id=50241592", "by": "ThunderSizzle", "date": "2026-10-26", "stance": "recommend", "quote": "one of the cheapest ways to get qwen3.8 27B Q6 with 100k+ context working", "excerpt": "If you already have a machine, throwing an r9700 in it is one of the cheapest ways to get qwen3.8 27B Q6 with 100k+ context working. Sadly, the r9700 has gone up in price, but it's still the cheapest route. I use It on an old x99 Linux desktop and it achieves 50 tps decode and 1000+ tps prefill.", "context": "runs it on an old x99 Linux desktop", "use_case": "local LLM on old desktop", "switched_from": "", "story": "IBM CEO says open weight models are now good and cheap on-prem"}, {"id": 800190672, "item_id": 50011917, "url": "https://news.ycombinator.com/item?id=50011917", "by": "roosterIllusi0n", "date": "2026-10-08", "stance": "recommend", "quote": "I have been using qwen3.8_27B_UD_Q3_K_XL. This does 46 tok/s on a 5080 with 16gb vram and it can pretty much do any systems work", "excerpt": "There are mixed weight models that make it work. I have been using qwen3.8_27B_UD_Q3_K_XL. This does 46 tok/s on a 5080 with 16gb vram and it can pretty much do any systems work because it can test to make sure it did it right.", "context": "", "use_case": "local systems coding", "switched_from": "", "story": "Sub-1-Bit LLM Compression via Latent Factorization"}, {"id": 797679840, "item_id": 49854990, "url": "https://news.ycombinator.com/item?id=49854990", "by": "praash", "date": "2026-09-26", "stance": "recommend", "quote": "Qwen 3.8 has been the first local model I've tried that has consistently given usable and impressive results.", "excerpt": "Qwen 3.8 27b-mlx running locally on Apple M4 48 GB. I only use local LLMs, mostly for getting technical suggestions and code snippet samples. Qwen 3.8 has been the first local model I've tried that has consistently given usable and impressive results.", "context": "only uses local LLMs on Apple M4 48 GB", "use_case": "local technical suggestions and code snippets", "switched_from": "", "story": "Ask HN: Which model do you use for work?"}, {"id": 800016320, "item_id": 50001020, "url": "https://news.ycombinator.com/item?id=50001020", "by": "latentsea", "date": "2026-10-08", "stance": "recommend", "quote": "I use Qwen3.8-27B as a daily driver", "excerpt": "I use Qwen3.8-27B as a daily driver, and for things I know will be quite hard I tend to get ChatGPT to do the planning. Works very well.", "context": "", "use_case": "daily coding driver", "switched_from": "", "story": "Meta and Microsoft take steps to reduce employee usage of Claude AI"}, {"id": 793086288, "item_id": 49567893, "url": "https://news.ycombinator.com/item?id=49567893", "by": "lrvick", "date": "2026-09-04", "stance": "recommend", "quote": "since 3.8 27b came out it is all I use", "excerpt": "Sure, I just add more cards as I need more throughput, and can combine up to four cards when I need 1M context on a smarter model, but in practice since 3.8 27b came out it is all I use. Also, unless your sessions are end to end encrypted to a secure enclave, then they are living in plain text -somewhere- and privacy policies tend to change when money is left on the table, if blackhats do not get to the data and sell", "context": "runs local GPUs", "use_case": "local LLM", "switched_from": "", "story": "Pause OpenAI Now"}]}, {"text": ", praising its coding ability for its size.", "cites": [{"id": 802092352, "item_id": 50130772, "url": "https://news.ycombinator.com/item?id=50130772", "by": "kamranjon", "date": "2026-10-17", "stance": "recommend", "quote": "Qwen is faster and seems to be much better at coding tasks.", "excerpt": "Try the Qwen 3.8 flash next model on DwarfStar - you will be very surprised. I\u2019ve been using the 4 bit quant almost exclusively after having DS v4 flash be my daily driver. Qwen is faster and seems to be much better at coding tasks.", "context": "", "use_case": "local coding", "switched_from": "DeepSeek V4 Flash", "story": "DeepSeek-v4.1-Flash-1017"}, {"id": 799288464, "item_id": 49955529, "url": "https://news.ycombinator.com/item?id=49955529", "by": "a11r", "date": "2026-10-04", "stance": "recommend", "quote": "We recently moved from 27B to Flash Next. The quality is superior for coding.", "excerpt": "We recently moved from 27B to Flash Next. The quality is superior for coding. Our workload is primarily well-defined coding tasks that need to be attempted a few times before the model gets it just right. FlashNext is also better at finding issues in generated code than Gemini 3.8 Flash.", "context": "team running a coding workload", "use_case": "well-defined coding tasks", "switched_from": "Qwen 3.8 27B", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 799009536, "item_id": 49938096, "url": "https://news.ycombinator.com/item?id=49938096", "by": "roosterIllusi0n", "date": "2026-10-02", "stance": "recommend", "quote": "qwen3.8 is more than capable for software dev.", "excerpt": "I am using qwen3.8-27B-UD-Q3_K_XL on a 5080 16gb card with 16gb of system ram at around 43-45tok/s with 65k context. I am using that model to set up containers on a proxmox server with two b70s and 96gb ram. The Q3 model implemented 10 different chat models, 2 image models, and 7 web based harnesses so I can compare. It can rewrite any part to improve it. qwen3.8 is more than capable for software dev. The gemma model", "context": "runs it locally on a 5080 16GB", "use_case": "local software development", "switched_from": "", "story": "Several vulnerabilities have been discovered in the Linux kernel"}, {"id": 799382960, "item_id": 49961435, "url": "https://news.ycombinator.com/item?id=49961435", "by": "Balinares", "date": "2026-10-05", "stance": "recommend", "quote": "3.8 is absolutely quite a bit better than 3.5.", "excerpt": "3.8 is absolutely quite a bit better than 3.5. My go-to quick benchmark is a bit like yours, implementing a simple but complete game in pygame. Most small models output code that's too buggy to be fixable. Qwen 3.8 wrote buggy code too, but the bugs were minor, and the game legitimately playable and fun. That said, I've not found it very reliable, especially at the low quant required to run on my hardware. It's very ", "context": "", "use_case": "local game code generation", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}]}, {"text": " The most common complaint is long reasoning that makes hard tasks slow", "cites": [{"id": 803774800, "item_id": 50235925, "url": "https://news.ycombinator.com/item?id=50235925", "by": "guregodevo", "date": "2026-10-25", "stance": "warn", "quote": "The main issue i found is that Qwen3.8-27B thinks a lot. On a hard problem the planner runs 32K tokens of reasoning and never writes the plan.", "excerpt": "I tried to compare with a workflow i built with this harness : https://github.com/guregodevo/memdoor/tree/main/examples/workflows/team-solve The main issue i found is that Qwen3.8-27B thinks a lot. On a hard problem the planner runs 32K tokens of reasoning and never writes the plan. Even with thinking off on Cerebras, a problem takes roughly 3 minutes and the 100 problems about 1.5 hours.", "context": "", "use_case": "workflow planning", "switched_from": "", "story": "Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness"}, {"id": 800546448, "item_id": 50034153, "url": "https://news.ycombinator.com/item?id=50034153", "by": "AustinDev", "date": "2026-10-10", "stance": "warn", "quote": "Quanted out Qwen 3.8 models are not frontier level. Have you used them for real work? They suck.", "excerpt": ">You can already performantly run frontier level models on systems with $1600 tier video cards. Quanted out Qwen 3.8 models are not frontier level. Have you used them for real work? They suck. They are so unsure of themselves due to the quant-lobotomy. They just endlessly loop in thinking even when they have the right decision and then output it when they get to the thinking cap. They score close in benchmarks becaus", "context": "", "use_case": "", "switched_from": "", "story": "100+ reactions to 100+ solutions"}, {"id": 794167632, "item_id": 49635477, "url": "https://news.ycombinator.com/item?id=49635477", "by": "anyfoo", "date": "2026-09-09", "stance": "warn", "quote": "Qwen3.8 27B does not seem to be good enough to implement the simulator with all its intricacies and tiny subtleties correctly", "excerpt": "You're probably partly right (though I use the same context size for all models, so there still is a difference in the model/quantization itself), but I definitely see it \"thinking itself into an extremely expensive and error-prone approach without backing out\". A common \"failure mode\" of my benchmark with You can already performantly run frontier level models on systems with $1600 tier video cards. Quanted out Qwen 3.8 models are not frontier level. Have you used them for real work? They suck. They are so unsure of themselves due to the quant-lobotomy. They just endlessly loop in thinking even when they have the right decision and then output it when they get to the thinking cap. They score close in benchmarks becaus", "context": "", "use_case": "", "switched_from": "", "story": "100+ reactions to 100+ solutions"}, {"id": 799382960, "item_id": 49961435, "url": "https://news.ycombinator.com/item?id=49961435", "by": "Balinares", "date": "2026-10-05", "stance": "recommend", "quote": "3.8 is absolutely quite a bit better than 3.5.", "excerpt": "3.8 is absolutely quite a bit better than 3.5. My go-to quick benchmark is a bit like yours, implementing a simple but complete game in pygame. Most small models output code that's too buggy to be fixable. Qwen 3.8 wrote buggy code too, but the bugs were minor, and the game legitimately playable and fun. That said, I've not found it very reliable, especially at the low quant required to run on my hardware. It's very ", "context": "", "use_case": "local game code generation", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}]}], "praise": [{"label": "Runs locally", "summary": "Commenters run it on single consumer GPUs, including 16GB and 24GB cards, and on Macs.", "quotes": [{"id": 800190672, "item_id": 50011917, "url": "https://news.ycombinator.com/item?id=50011917", "by": "roosterIllusi0n", "date": "2026-10-08", "stance": "recommend", "quote": "I have been using qwen3.8_27B_UD_Q3_K_XL. This does 46 tok/s on a 5080 with 16gb vram and it can pretty much do any systems work", "excerpt": "There are mixed weight models that make it work. I have been using qwen3.8_27B_UD_Q3_K_XL. This does 46 tok/s on a 5080 with 16gb vram and it can pretty much do any systems work because it can test to make sure it did it right.", "context": "", "use_case": "local systems coding", "switched_from": "", "story": "Sub-1-Bit LLM Compression via Latent Factorization"}, {"id": 802696816, "item_id": 50168551, "url": "https://news.ycombinator.com/item?id=50168551", "by": "nacs", "date": "2026-10-20", "stance": "recommend", "quote": "I use it at 60+ tok/second on a single 24GB card.", "excerpt": "With Strata and specialized engines, Qwen Flash Next runs even on 16GB cards. I use it at 60+ tok/second on a single 24GB card.", "context": "runs local models", "use_case": "local inference", "switched_from": "", "story": "MiniMax M3.1 Flash"}, {"id": 792240320, "item_id": 49515020, "url": "https://news.ycombinator.com/item?id=49515020", "by": "jckahn", "date": "2026-08-31", "stance": "recommend", "quote": "I run Qwen 3.8 27B just fine on my Mac mini M4 24GB.", "excerpt": "I run Qwen 3.8 27B just fine on my Mac mini M4 24GB. I use Unsloth's Q3 XXS with 128k context. It successfully completes long horizon tasks with OpenCode.", "context": "", "use_case": "local long-horizon coding tasks", "switched_from": "", "story": "Apple caught off guard by AI demand for Mac Mini and Mac Studio"}]}, {"label": "Coding ability", "summary": "Users find it strong for software development, with some saying it beats DeepSeek V4 Flash at coding.", "quotes": [{"id": 802092352, "item_id": 50130772, "url": "https://news.ycombinator.com/item?id=50130772", "by": "kamranjon", "date": "2026-10-17", "stance": "recommend", "quote": "Qwen is faster and seems to be much better at coding tasks.", "excerpt": "Try the Qwen 3.8 flash next model on DwarfStar - you will be very surprised. I\u2019ve been using the 4 bit quant almost exclusively after having DS v4 flash be my daily driver. Qwen is faster and seems to be much better at coding tasks.", "context": "", "use_case": "local coding", "switched_from": "DeepSeek V4 Flash", "story": "DeepSeek-v4.1-Flash-1017"}, {"id": 799288464, "item_id": 49955529, "url": "https://news.ycombinator.com/item?id=49955529", "by": "a11r", "date": "2026-10-04", "stance": "recommend", "quote": "We recently moved from 27B to Flash Next. The quality is superior for coding.", "excerpt": "We recently moved from 27B to Flash Next. The quality is superior for coding. Our workload is primarily well-defined coding tasks that need to be attempted a few times before the model gets it just right. FlashNext is also better at finding issues in generated code than Gemini 3.8 Flash.", "context": "team running a coding workload", "use_case": "well-defined coding tasks", "switched_from": "Qwen 3.8 27B", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 799009536, "item_id": 49938096, "url": "https://news.ycombinator.com/item?id=49938096", "by": "roosterIllusi0n", "date": "2026-10-02", "stance": "recommend", "quote": "qwen3.8 is more than capable for software dev.", "excerpt": "I am using qwen3.8-27B-UD-Q3_K_XL on a 5080 16gb card with 16gb of system ram at around 43-45tok/s with 65k context. I am using that model to set up containers on a proxmox server with two b70s and 96gb ram. The Q3 model implemented 10 different chat models, 2 image models, and 7 web based harnesses so I can compare. It can rewrite any part to improve it. qwen3.8 is more than capable for software dev. The gemma model", "context": "runs it locally on a 5080 16GB", "use_case": "local software development", "switched_from": "", "story": "Several vulnerabilities have been discovered in the Linux kernel"}]}, {"label": "Daily driver", "summary": "Several commenters say it is the main or only model they use day to day.", "quotes": [{"id": 800016320, "item_id": 50001020, "url": "https://news.ycombinator.com/item?id=50001020", "by": "latentsea", "date": "2026-10-08", "stance": "recommend", "quote": "I use Qwen3.8-27B as a daily driver", "excerpt": "I use Qwen3.8-27B as a daily driver, and for things I know will be quite hard I tend to get ChatGPT to do the planning. Works very well.", "context": "", "use_case": "daily coding driver", "switched_from": "", "story": "Meta and Microsoft take steps to reduce employee usage of Claude AI"}, {"id": 793086288, "item_id": 49567893, "url": "https://news.ycombinator.com/item?id=49567893", "by": "lrvick", "date": "2026-09-04", "stance": "recommend", "quote": "since 3.8 27b came out it is all I use", "excerpt": "Sure, I just add more cards as I need more throughput, and can combine up to four cards when I need 1M context on a smarter model, but in practice since 3.8 27b came out it is all I use. Also, unless your sessions are end to end encrypted to a secure enclave, then they are living in plain text -somewhere- and privacy policies tend to change when money is left on the table, if blackhats do not get to the data and sell", "context": "runs local GPUs", "use_case": "local LLM", "switched_from": "", "story": "Pause OpenAI Now"}, {"id": 798446128, "item_id": 49902883, "url": "https://news.ycombinator.com/item?id=49902883", "by": "latentsea", "date": "2026-09-30", "stance": "recommend", "quote": "Qwen3.8-27B is a huge step up from Qwen3.6-27B.", "excerpt": "Qwen3.8-27B is a huge step up from Qwen3.6-27B. That release only happened relatively recently but that felt like the 'Opus 4.5' release turning point that SOTA models experienced back when that came out. It was the first time I felt like local models are actually good enough to use as daily drivers now. So, it was only after that point that I switched. Qwen3.8-Flash-Next is better still if you can run fast enough. I", "context": "runs local models at home", "use_case": "local daily driver", "switched_from": "", "story": "GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price"}]}, {"label": "Fast throughput", "summary": "Users report high token rates on local hardware, especially with Flash Next and multiple streams.", "quotes": [{"id": 803825184, "item_id": 50239074, "url": "https://news.ycombinator.com/item?id=50239074", "by": "bitexploder", "date": "2026-10-26", "stance": "recommend", "quote": "My server can run Qwen Flash Next at 500 t/s aggregate (about 200 t/s single user). It's pretty great.", "excerpt": "I mean, I just have 4xV100S GPU and a Macbook Pro w/48GB of RAM and I find it kind of amazing. My server can run Qwen Flash Next at 500 t/s aggregate (about 200 t/s single user). It's pretty great. I use it for local coding and all my worker agents. It is about as reliable as Sonnet 5 for agentic coding. I make 3D models, keep the home network and server updated, fully built ansible files. You do NOT need 20K in hard", "context": "runs 4xV100S GPU server at home", "use_case": "local coding and worker agents", "switched_from": "", "story": "IBM CEO says open weight models are now good and cheap on-prem"}, {"id": 793846832, "item_id": 49615427, "url": "https://news.ycombinator.com/item?id=49615427", "by": "skolos", "date": "2026-09-08", "stance": "recommend", "quote": "qwen3.8 27b is different.", "excerpt": "On many models that I tested in past context quantization had very bad effect on model performance. However qwen3.8 27b is different. I'm now running NVFP4 quantized both weight and cache on my RTX5090 and getting excellent results: 264k cache allocated for pool, 10k tok/s prompt processing, 200 tok/s generation for single stream, or 801 tok/s generation for 8 concurrent streams. Also have about 2Gb vram left for use", "context": "runs it on an RTX 5090", "use_case": "local coding agents", "switched_from": "", "story": "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses"}, {"id": 802549168, "item_id": 50159323, "url": "https://news.ycombinator.com/item?id=50159323", "by": "bitexploder", "date": "2026-10-19", "stance": "recommend", "quote": "I am just so impressed with Qwen Flash Next right now. I have been using it locally and it's just such a great little model to work with.", "excerpt": "I know I mention it way too often right now, but I am just so impressed with Qwen Flash Next right now. I have been using it locally and it's just such a great little model to work with. It makes mistakes and has its quirks, but in general, it just gets things done and its default manner is just right. It is not overly proactive or making messes for me, even compared to DeepSeek Flash 4.1. I really like DS Flash too,", "context": "runs it locally on DDR4 and old data center GPUs", "use_case": "local coding model", "switched_from": "", "story": "MiniMax M3.1 Flash"}]}, {"label": "Near-frontier quality", "summary": "Some compare it to Sonnet or Opus for coding at far lower cost, though slower.", "quotes": [{"id": 798966688, "item_id": 49935418, "url": "https://news.ycombinator.com/item?id=49935418", "by": "Roark66", "date": "2026-10-02", "stance": "recommend", "quote": "I've been running it locally since for real programming and I consider it pretty much the same as opus 4.6 in coding ability", "excerpt": "I wonder if August of this year will be remembered like that too. The very first \"opus like\" local AI model came out this August (Qwen 3.8 Flash Next). I've been running it locally since for real programming and I consider it pretty much the same as opus 4.6 in coding ability (it lacks a bit in the factual knowledge area). It even exceeds opus on some tasks. I've tried every hyped \"open source\" model before and all i", "context": "runs it on 6 RTX 3090s", "use_case": "local programming", "switched_from": "Claude Code", "story": "Several vulnerabilities have been discovered in the Linux kernel"}, {"id": 797400370, "item_id": 49837523, "url": "https://news.ycombinator.com/item?id=49837523", "by": "sznio", "date": "2026-09-24", "stance": "recommend", "quote": "decently replaces Sonnet, albeit not as fast", "excerpt": "Nothing, because current Haiku is useless. But I use Qwen3.6 35B-A3B and it's both much faster and much more capable than Haiku. And Qwen3.8 27B can do so much more, decently replaces Sonnet, albeit not as fast. Luna doesn't feel more capable nor faster than Qwen3.8. I think Anthropic could surprise us with some more innovation in this niche.", "context": "", "use_case": "", "switched_from": "", "story": "Claude Opus 5.5"}, {"id": 798868402, "item_id": 49929275, "url": "https://news.ycombinator.com/item?id=49929275", "by": "pettijohn", "date": "2026-10-02", "stance": "recommend", "quote": "achieving flagship quality without the price tag (albeit slower)", "excerpt": "I use PI WEB (https://pi-web.dev, there's more than one with that name) to orchestrate remotely. Pi lets me run local models (currently Qwen 3.8 Flash Next on Strix Halo 128GB) and flagship models (with subscription auth, not API pricing) side by side. I typically ask GPT 6.1 Sol to review requirements and then spawn a subsession with Qwen, review Qwen's work, ask Qwen to fix. About 95% success with a single pass lik", "context": "runs local models on Strix Halo 128GB", "use_case": "local implementation model", "switched_from": "", "story": "Pi 1.0"}]}, {"label": "Long context", "summary": "Commenters report working sessions up to 200k+ context without noticeable degradation.", "quotes": [{"id": 793846832, "item_id": 49615427, "url": "https://news.ycombinator.com/item?id=49615427", "by": "skolos", "date": "2026-09-08", "stance": "recommend", "quote": "qwen3.8 27b is different.", "excerpt": "On many models that I tested in past context quantization had very bad effect on model performance. However qwen3.8 27b is different. I'm now running NVFP4 quantized both weight and cache on my RTX5090 and getting excellent results: 264k cache allocated for pool, 10k tok/s prompt processing, 200 tok/s generation for single stream, or 801 tok/s generation for 8 concurrent streams. Also have about 2Gb vram left for use", "context": "runs it on an RTX 5090", "use_case": "local coding agents", "switched_from": "", "story": "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses"}, {"id": 798939809, "item_id": 49933738, "url": "https://news.ycombinator.com/item?id=49933738", "by": "tingletech", "date": "2026-10-02", "stance": "recommend", "quote": "I use it with a local Qwen3.8 27B on llama-server", "excerpt": "I use it with a local Qwen3.8 27B on llama-server, but I don't know what it means for it exceed any window. you can use /slots on the llama-server if you want to get more up-to-date details on session token use. I always run models in their default context, which is 262144 for Qwen3.8 27B. I've run sessions in hax where it hit that cap multiple times and compressed the context down to 15% and continued with no proble", "context": "", "use_case": "local model for agent sessions", "switched_from": "", "story": "Pi 1.0"}, {"id": 803865472, "item_id": 50241592, "url": "https://news.ycombinator.com/item?id=50241592", "by": "ThunderSizzle", "date": "2026-10-26", "stance": "recommend", "quote": "one of the cheapest ways to get qwen3.8 27B Q6 with 100k+ context working", "excerpt": "If you already have a machine, throwing an r9700 in it is one of the cheapest ways to get qwen3.8 27B Q6 with 100k+ context working. Sadly, the r9700 has gone up in price, but it's still the cheapest route. I use It on an old x99 Linux desktop and it achieves 50 tps decode and 1000+ tps prefill.", "context": "runs it on an old x99 Linux desktop", "use_case": "local LLM on old desktop", "switched_from": "", "story": "IBM CEO says open weight models are now good and cheap on-prem"}]}], "caveats": [{"label": "Overthinking", "summary": "It can reason for very long stretches, sometimes never finishing or taking much time and many tokens.", "quotes": [{"id": 803774800, "item_id": 50235925, "url": "https://news.ycombinator.com/item?id=50235925", "by": "guregodevo", "date": "2026-10-25", "stance": "warn", "quote": "The main issue i found is that Qwen3.8-27B thinks a lot. On a hard problem the planner runs 32K tokens of reasoning and never writes the plan.", "excerpt": "I tried to compare with a workflow i built with this harness : https://github.com/guregodevo/memdoor/tree/main/examples/workflows/team-solve The main issue i found is that Qwen3.8-27B thinks a lot. On a hard problem the planner runs 32K tokens of reasoning and never writes the plan. Even with thinking off on Cerebras, a problem takes roughly 3 minutes and the 100 problems about 1.5 hours.", "context": "", "use_case": "workflow planning", "switched_from": "", "story": "Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness"}, {"id": 800546448, "item_id": 50034153, "url": "https://news.ycombinator.com/item?id=50034153", "by": "AustinDev", "date": "2026-10-10", "stance": "warn", "quote": "Quanted out Qwen 3.8 models are not frontier level. Have you used them for real work? They suck.", "excerpt": ">You can already performantly run frontier level models on systems with $1600 tier video cards. Quanted out Qwen 3.8 models are not frontier level. Have you used them for real work? They suck. They are so unsure of themselves due to the quant-lobotomy. They just endlessly loop in thinking even when they have the right decision and then output it when they get to the thinking cap. They score close in benchmarks becaus", "context": "", "use_case": "", "switched_from": "", "story": "100+ reactions to 100+ solutions"}, {"id": 791244560, "item_id": 49452785, "url": "https://news.ycombinator.com/item?id=49452785", "by": "skohan", "date": "2026-08-26", "stance": "recommend", "quote": "It's the first small local model I've felt like I can do real work with.", "excerpt": "I've had the exact opposite experience. I've been using 3.8 for my daily driver since last week, and I've gradually been giving it more and more complex tasks as it continues to deliver high quality results. Now I am basically handing off large complex features, and 3.8 is doing the planning, task breakdown, implementation and review with just a few notes from my side. The tradeoff is time (especially on RDMA4 hardwa", "context": "", "use_case": "local agentic coding", "switched_from": "", "story": "GLM-5.3-Flash"}]}, {"label": "Quantization losses", "summary": "Low-bit quants are described as less reliable and prone to looping or inconsistent results.", "quotes": [{"id": 800546448, "item_id": 50034153, "url": "https://news.ycombinator.com/item?id=50034153", "by": "AustinDev", "date": "2026-10-10", "stance": "warn", "quote": "Quanted out Qwen 3.8 models are not frontier level. Have you used them for real work? They suck.", "excerpt": ">You can already performantly run frontier level models on systems with $1600 tier video cards. Quanted out Qwen 3.8 models are not frontier level. Have you used them for real work? They suck. They are so unsure of themselves due to the quant-lobotomy. They just endlessly loop in thinking even when they have the right decision and then output it when they get to the thinking cap. They score close in benchmarks becaus", "context": "", "use_case": "", "switched_from": "", "story": "100+ reactions to 100+ solutions"}, {"id": 799382960, "item_id": 49961435, "url": "https://news.ycombinator.com/item?id=49961435", "by": "Balinares", "date": "2026-10-05", "stance": "recommend", "quote": "3.8 is absolutely quite a bit better than 3.5.", "excerpt": "3.8 is absolutely quite a bit better than 3.5. My go-to quick benchmark is a bit like yours, implementing a simple but complete game in pygame. Most small models output code that's too buggy to be fixable. Qwen 3.8 wrote buggy code too, but the bugs were minor, and the game legitimately playable and fun. That said, I've not found it very reliable, especially at the low quant required to run on my hardware. It's very ", "context": "", "use_case": "local game code generation", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}]}, {"label": "Slow on Macs", "summary": "Some Apple Silicon users found the 27B model too slow to use comfortably.", "quotes": [{"id": 795681440, "item_id": 49730090, "url": "https://news.ycombinator.com/item?id=49730090", "by": "bellowsgulch", "date": "2026-09-16", "stance": "warn", "quote": "It's just too slow of a model.", "excerpt": "I so wanted to be able to use Qwen3.8-27B locally on my fully loaded M1 Max 64GB when spider-mario recommended I use MTPLX[1], but I still found the general MLX MTP acceleration to be so poor, I'd rather just use cheap tokens from OpenCode Zen and Go. It's just too slow of a model. I know, I know there's new hardware, but my business paid like 4-5 grand for this MBP at the time, and I just don't feel the need to pay ", "context": "business owner with an M1 Max 64GB MacBook Pro", "use_case": "local inference", "switched_from": "", "story": "Show HN: Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh"}, {"id": 791704417, "item_id": 49481526, "url": "https://news.ycombinator.com/item?id=49481526", "by": "pwython", "date": "2026-08-28", "stance": "warn", "quote": "Qwen3.8 wasn't fun to use on my 64GB M4 Max either", "excerpt": "Yea Qwen3.8 wasn't fun to use on my 64GB M4 Max either (better than these numbers though), so my new daily driver is Ornith-1.5-35B-A3B-MLX-4bit. I recommend giving that a whirl if you're on similar hardware, it's definitely better than Qwen3.6 35b-a3b which was my go-to before. https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B", "context": "64GB M4 Max user", "use_case": "local LLM on Mac", "switched_from": "", "story": "Run Qwen3.8 27B locally: real numbers from my Mac Studio"}]}, {"label": "Nuanced tasks", "summary": "Some find it falls short of frontier models on complex or subtle work.", "quotes": [{"id": 793826145, "item_id": 49614134, "url": "https://news.ycombinator.com/item?id=49614134", "by": "TomBombadildoze", "date": "2026-09-08", "stance": "warn", "quote": "I have found qwen to be absolutely useless at anything but very specific, surgical code changes.", "excerpt": "Beats Opus 4.6 at what exactly? It certainly isn't code. I use a combination of a Claude Max subscription and local inference, including qwen3.8-27b, 4bit. I have found qwen to be absolutely useless at anything but very specific, surgical code changes. In my experience, for anything even remotely nuanced, a frontier model is required.", "context": "", "use_case": "local coding", "switched_from": "", "story": "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses"}, {"id": 794167632, "item_id": 49635477, "url": "https://news.ycombinator.com/item?id=49635477", "by": "anyfoo", "date": "2026-09-09", "stance": "warn", "quote": "Qwen3.8 27B does not seem to be good enough to implement the simulator with all its intricacies and tiny subtleties correctly", "excerpt": "You're probably partly right (though I use the same context size for all models, so there still is a difference in the model/quantization itself), but I definitely see it \"thinking itself into an extremely expensive and error-prone approach without backing out\". A common \"failure mode\" of my benchmark with /task.md - the problem statement /plan.md - the manager\u2019s overarching plan /tasks.json - a debug mirror of the task list /notes.md - curated ideas / findings / partial proofs /solution.py - the current solution Then they describe h", "context": "runs it in a container", "use_case": "agentic coding with plan files", "switched_from": "", "story": "Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness"}, {"id": 803556580, "item_id": 50222286, "url": "https://news.ycombinator.com/item?id=50222286", "by": "potsandpans", "date": "2026-10-24", "stance": "neutral", "quote": "Qwen 3.8 27b on my 5090, barely acceptable, thinks too long, sometimes freaks out.", "excerpt": "I recently just crossed the threshold for 100% decompilation of a PlayStation game. I used a lot of models in the process, and I feel like this is a good demonstration the \"sauce\" that opus has. The harness I've built for it is pretty straight forward: split binary into individual functions, give the agent a bunch of small tools to diff output with underlying binary, make them wrote clean c until there's a certified ", "context": "decompiling a PlayStation game", "use_case": "local decompilation agent", "switched_from": "", "story": "Kimi K3.1 scores 55 on Artificial Analysis, placed on the Pareto line"}, {"id": 802696816, "item_id": 50168551, "url": "https://news.ycombinator.com/item?id=50168551", "by": "nacs", "date": "2026-10-20", "stance": "recommend", "quote": "I use it at 60+ tok/second on a single 24GB card.", "excerpt": "With Strata and specialized engines, Qwen Flash Next runs even on 16GB cards. I use it at 60+ tok/second on a single 24GB card.", "context": "runs local models", "use_case": "local inference", "switched_from": "", "story": "MiniMax M3.1 Flash"}, {"id": 802549168, "item_id": 50159323, "url": "https://news.ycombinator.com/item?id=50159323", "by": "bitexploder", "date": "2026-10-19", "stance": "recommend", "quote": "I am just so impressed with Qwen Flash Next right now. I have been using it locally and it's just such a great little model to work with.", "excerpt": "I know I mention it way too often right now, but I am just so impressed with Qwen Flash Next right now. I have been using it locally and it's just such a great little model to work with. It makes mistakes and has its quirks, but in general, it just gets things done and its default manner is just right. It is not overly proactive or making messes for me, even compared to DeepSeek Flash 4.1. I really like DS Flash too,", "context": "runs it locally on DDR4 and old data center GPUs", "use_case": "local coding model", "switched_from": "", "story": "MiniMax M3.1 Flash"}, {"id": 802501104, "item_id": 50156319, "url": "https://news.ycombinator.com/item?id=50156319", "by": "segmondy", "date": "2026-10-19", "stance": "recommend", "quote": "Both were great, but surprisingly, Q3.8FN had the better result.", "excerpt": "Too early to call, I ran it through the weekend and compared with Qwen3.8-Flash-Next, DSv41F-1017 at FP8 and Q3.8FN at Q8. Both were great, but surprisingly, Q3.8FN had the better result. I was modelling and generating simulation code for a complex system. I used both because I had a bunch of images that needed to be used for input to steer the output. I think these models have strengths and it's best to experiment a", "context": "", "use_case": "simulation code generation", "switched_from": "", "story": "DeepSeek V4.1 Flash 1017 (Max)"}, {"id": 802092352, "item_id": 50130772, "url": "https://news.ycombinator.com/item?id=50130772", "by": "kamranjon", "date": "2026-10-17", "stance": "recommend", "quote": "Qwen is faster and seems to be much better at coding tasks.", "excerpt": "Try the Qwen 3.8 flash next model on DwarfStar - you will be very surprised. I\u2019ve been using the 4 bit quant almost exclusively after having DS v4 flash be my daily driver. Qwen is faster and seems to be much better at coding tasks.", "context": "", "use_case": "local coding", "switched_from": "DeepSeek V4 Flash", "story": "DeepSeek-v4.1-Flash-1017"}, {"id": 801962513, "item_id": 50122657, "url": "https://news.ycombinator.com/item?id=50122657", "by": "colingauvin", "date": "2026-10-16", "stance": "compare", "quote": "I switched to Qwen for image input", "excerpt": "I was running DS731 until Qwen 3.8 Flash Next (so what, like a couple weeks? The pace has been torrid). I switched to Qwen for image input, then GLM 5.3 Flash. At the time, Qwen felt like a minor step down from DS731, GLM 5.3 Flash feels like a major step up from Qwen, so I'd say slightly better than DS731. I realize that's not exactly what the benchmarks say, but DS731 definitely flailed around on some of the more c", "context": "", "use_case": "image input", "switched_from": "DeepSeek 0731", "story": "Muse Spark Open"}, {"id": 801896961, "item_id": 50118560, "url": "https://news.ycombinator.com/item?id=50118560", "by": "blinkingled", "date": "2026-10-16", "stance": "compare", "quote": "Yes 27B - not so much of a surprise but I was hoping for a usable coding model that's local and fast", "excerpt": "Yes 27B - not so much of a surprise but I was hoping for a usable coding model that's local and fast - this is fast but maybe not usable. I used Gemma4 26B before and it wasn't good in my testing.", "context": "", "use_case": "local coding model", "switched_from": "", "story": "LFM2.5-24B-A2B: Built for fast on-device agents"}, {"id": 800702929, "item_id": 50043933, "url": "https://news.ycombinator.com/item?id=50043933", "by": "Majromax", "date": "2026-10-11", "stance": "neutral", "quote": "locally hosted Qwen-3.8-27b or Qwen-Flash-Next", "excerpt": "> One thing that's different about Sinai is that it keeps the code author and reviewer in separate agents with separate context. At the time we built it, we suspected that this would lead to better outcomes, but again, we haven't validated that it does. I use a similar process for a local agent swarm approach (locally hosted Qwen-3.8-27b or Qwen-Flash-Next), used so far for personal-grade projects. Through tool and p", "context": "", "use_case": "local agent swarm", "switched_from": "", "story": "Five months treating bugs like patients and coding agents like a medical team"}, {"id": 800702930, "item_id": 50043933, "url": "https://news.ycombinator.com/item?id=50043933", "by": "Majromax", "date": "2026-10-11", "stance": "neutral", "quote": "locally hosted Qwen-3.8-27b or Qwen-Flash-Next", "excerpt": "> One thing that's different about Sinai is that it keeps the code author and reviewer in separate agents with separate context. At the time we built it, we suspected that this would lead to better outcomes, but again, we haven't validated that it does. I use a similar process for a local agent swarm approach (locally hosted Qwen-3.8-27b or Qwen-Flash-Next), used so far for personal-grade projects. Through tool and p", "context": "", "use_case": "local agent swarm", "switched_from": "", "story": "Five months treating bugs like patients and coding agents like a medical team"}, {"id": 800546448, "item_id": 50034153, "url": "https://news.ycombinator.com/item?id=50034153", "by": "AustinDev", "date": "2026-10-10", "stance": "warn", "quote": "Quanted out Qwen 3.8 models are not frontier level. Have you used them for real work? They suck.", "excerpt": ">You can already performantly run frontier level models on systems with $1600 tier video cards. Quanted out Qwen 3.8 models are not frontier level. Have you used them for real work? They suck. They are so unsure of themselves due to the quant-lobotomy. They just endlessly loop in thinking even when they have the right decision and then output it when they get to the thinking cap. They score close in benchmarks becaus", "context": "", "use_case": "", "switched_from": "", "story": "100+ reactions to 100+ solutions"}, {"id": 800242066, "item_id": 50015129, "url": "https://news.ycombinator.com/item?id=50015129", "by": "lisplist", "date": "2026-10-09", "stance": "recommend", "quote": "if you're VRAM constrained, take a look at GLM 5.3 Flash or Qwen 3.8 Flash Next", "excerpt": "DeepSeek V4.1 Flash is mixed MXFP4/MXFP8 so all but the INT4 calculation is wrong here and that's still wrong because you can run it on < 400GB VRAM. The n-gram table is MXFP8, but can be offloaded to RAM or disk without too much of a performance hit. Really, you could probably cram it onto < 300GB VRAM if you're willing to apply a small quant to certain parts of the model considering how little VRAM is dedicated to ", "context": "", "use_case": "VRAM-constrained local inference", "switched_from": "", "story": "Why isn't the industry freaking out about DeepSeek 4.1 Flash?"}, {"id": 800190672, "item_id": 50011917, "url": "https://news.ycombinator.com/item?id=50011917", "by": "roosterIllusi0n", "date": "2026-10-08", "stance": "recommend", "quote": "I have been using qwen3.8_27B_UD_Q3_K_XL. This does 46 tok/s on a 5080 with 16gb vram and it can pretty much do any systems work", "excerpt": "There are mixed weight models that make it work. I have been using qwen3.8_27B_UD_Q3_K_XL. This does 46 tok/s on a 5080 with 16gb vram and it can pretty much do any systems work because it can test to make sure it did it right.", "context": "", "use_case": "local systems coding", "switched_from": "", "story": "Sub-1-Bit LLM Compression via Latent Factorization"}, {"id": 800188178, "item_id": 50011761, "url": "https://news.ycombinator.com/item?id=50011761", "by": "krees", "date": "2026-10-08", "stance": "neutral", "quote": "Qwen 3.8 27b / Gemma 12b", "excerpt": "Mac M3, oMLX, pi. 64GB RAM. Qwen 3.8 27b / Gemma 12b", "context": "", "use_case": "local model for agents", "switched_from": "", "story": "Ask HN: Got a Mac for running agents? what's your setup?"}, {"id": 800172211, "item_id": 50010763, "url": "https://news.ycombinator.com/item?id=50010763", "by": "jwr", "date": "2026-10-08", "stance": "neutral", "quote": "feeding the results through local Qwen 3.8 running under MTPLX", "excerpt": "People keep praising Parakeet, but I've found it to be worse than Whisper Large. Yes, it is much, much faster and smaller, but accuracy matters a lot if you are to use dictation regularly and seriously. I ended up having AI optimize Whisper Large and create a plugin for TypeWhisper, and that's what I use (feeding the results through local Qwen 3.8 running under MTPLX).", "context": "", "use_case": "cleaning up dictation", "switched_from": "", "story": "Whistle: Speech to Text in 16.9 MB"}, {"id": 800016320, "item_id": 50001020, "url": "https://news.ycombinator.com/item?id=50001020", "by": "latentsea", "date": "2026-10-08", "stance": "recommend", "quote": "I use Qwen3.8-27B as a daily driver", "excerpt": "I use Qwen3.8-27B as a daily driver, and for things I know will be quite hard I tend to get ChatGPT to do the planning. Works very well.", "context": "", "use_case": "daily coding driver", "switched_from": "", "story": "Meta and Microsoft take steps to reduce employee usage of Claude AI"}, {"id": 799810304, "item_id": 49988144, "url": "https://news.ycombinator.com/item?id=49988144", "by": "colingauvin", "date": "2026-10-07", "stance": "recommend", "quote": "My tests with Qwen 3.8 have convinced me it's completely fine general purpose", "excerpt": "The author's configuration is..... Interesting. I don't disagree with the title. This spring folks were saying \"I use Qwen 3.6 27B (or 35 A3B) as a daily driver\". They either had low standards or were lying. The models were good enough technically, they capable enough, that you could see the potential, but the coherence was really not great compared to frontier. They'd lose track of what they were doing, mix things u", "context": "runs models on 2x Sparks", "use_case": "local general-purpose model", "switched_from": "", "story": "Local AI is real now. And it's blowing my mind"}, {"id": 799704512, "item_id": 49981532, "url": "https://news.ycombinator.com/item?id=49981532", "by": "ineptech", "date": "2026-10-06", "stance": "recommend", "quote": "I use Qwen 3.8 26b at home and find it very useful for some tasks", "excerpt": "I follow the same approach (finding LLMs useful but only running open models on my own hardware) but I've been limited by also not letting them into my email and being pretty restrictive about their internet access. Would you mind sharing what stack and agent harness you use, since your other comments make it clear you're highly restrictive about what kinds of tech you run at home and I assume you vetted it thoroughl", "context": "runs open models on own hardware", "use_case": "local coding assistant", "switched_from": "", "story": "Meta\u2019s Muse is an adorable privacy and security dumpster fire"}, {"id": 799665717, "item_id": 49979107, "url": "https://news.ycombinator.com/item?id=49979107", "by": "segmondy", "date": "2026-10-06", "stance": "neutral", "quote": "I use MiMov2.6Pro, DeepSeekv4.1Flash, GLM5.3, Hy4, Qwen3.8 and KimiK3 at home.", "excerpt": "I use Opus 5.5 at work. I use MiMov2.6Pro, DeepSeekv4.1Flash, GLM5.3, Hy4, Qwen3.8 and KimiK3 at home. Opus5.5 is not a game changer.", "context": "", "use_case": "home use", "switched_from": "", "story": "Mistral Large 4"}, {"id": 799436480, "item_id": 49964780, "url": "https://news.ycombinator.com/item?id=49964780", "by": "bitexploder", "date": "2026-10-05", "stance": "recommend", "quote": "this model is the first local model that does a good job. It is kind of amazing.", "excerpt": "Real world: I use QFN38 to build 3d models. I use build123d + blender and this model is the first local model that does a good job. It is kind of amazing. This model is as good as DSV 4.1 on most tasks and I think it only loses when specific knowledge matters.", "context": "", "use_case": "building 3D models", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 799382960, "item_id": 49961435, "url": "https://news.ycombinator.com/item?id=49961435", "by": "Balinares", "date": "2026-10-05", "stance": "recommend", "quote": "3.8 is absolutely quite a bit better than 3.5.", "excerpt": "3.8 is absolutely quite a bit better than 3.5. My go-to quick benchmark is a bit like yours, implementing a simple but complete game in pygame. Most small models output code that's too buggy to be fixable. Qwen 3.8 wrote buggy code too, but the bugs were minor, and the game legitimately playable and fun. That said, I've not found it very reliable, especially at the low quant required to run on my hardware. It's very ", "context": "", "use_case": "local game code generation", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 799288464, "item_id": 49955529, "url": "https://news.ycombinator.com/item?id=49955529", "by": "a11r", "date": "2026-10-04", "stance": "recommend", "quote": "We recently moved from 27B to Flash Next. The quality is superior for coding.", "excerpt": "We recently moved from 27B to Flash Next. The quality is superior for coding. Our workload is primarily well-defined coding tasks that need to be attempted a few times before the model gets it just right. FlashNext is also better at finding issues in generated code than Gemini 3.8 Flash.", "context": "team running a coding workload", "use_case": "well-defined coding tasks", "switched_from": "Qwen 3.8 27B", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 799264416, "item_id": 49954026, "url": "https://news.ycombinator.com/item?id=49954026", "by": "roscas", "date": "2026-10-04", "stance": "recommend", "quote": "I use it with the browser that has a great dashboard and with hermes agent and that it really makes this amazing.", "excerpt": "Coder version with 30t/sec on a Ryzen 3600x with 48GB of RAM with a nvidia 3080. This is not a very fast desktop. Memory speed is around 2000mhz only. My SSD is some of the worst SSD I've seen and 3080 had its days of glory. I still have code, chromium, librewolf and many other programs running. I have video streams running while I also watch tv and many times youtube videos. I use it with the browser that has a grea", "context": "Ryzen 3600x, 48GB RAM, RTX 3080 desktop", "use_case": "local coding model", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 799254753, "item_id": 49953422, "url": "https://news.ycombinator.com/item?id=49953422", "by": "cyanydeez", "date": "2026-10-04", "stance": "neutral", "quote": "I'm working on local models; qwen3.8-Flash-Next seems to be passed to rubicon.", "excerpt": "I'm working on local models; qwen3.8-Flash-Next seems to be passed to rubicon. I'm working on javascript, and now I'm only doing this via typescript. I've successfully got this going: 1. Design a feature in plain language with the coding agen (opencode) and write a document for it. 2. Restart the context (or I use /compact to flush any errant details) 3. Pull the new plan into context and ask the agent to revise it f", "context": "", "use_case": "local coding", "switched_from": "", "story": "Agents don't need memory, they need documentation"}, {"id": 799111936, "item_id": 49944496, "url": "https://news.ycombinator.com/item?id=49944496", "by": "mark_l_watson", "date": "2026-10-03", "stance": "recommend", "quote": "the results are so very much excellent enough for about 60% of my work", "excerpt": "It was inevitable that the providers would stop giving away free cookies and milkshakes. I pay Google $5/month for Chat and my wife gets to share that: a fantastic deal. I even get a short weekly (one each) opus and gemini vibe coding sessions with agy. I pay for inference: I paid $3300 for a Mac mini 64G M5-Pro which runs Qwen-3.8-Flash-Next (3 bit quant) with pi-dev and the results are so very much excellent enough", "context": "", "use_case": "local inference for coding work", "switched_from": "", "story": "Gemini ending free use of Flash and Pro models"}, {"id": 799032352, "item_id": 49939522, "url": "https://news.ycombinator.com/item?id=49939522", "by": "sleight42", "date": "2026-10-02", "stance": "recommend", "quote": "That's good enough for me for development.", "excerpt": "I don't do that. I use my 2020 top end gaming PC with its 3090. I just ordered 128GB RAM for it. I'll use a single-chat/slot runtime like Strata or Freetoken. And then I'll be able to run either a blazing fast 8-bit Qwen 3.8 27b or a 4 or 5 bit Qwen 3.8 Flash Next. That's good enough for me for development. And then I'll have my current gaming rig with less RAM but better CPU and GPU run a reasonable smart tool agent", "context": "has a 2020 gaming PC with a 3090", "use_case": "local development model", "switched_from": "", "story": "Dots: Always-on agents"}, {"id": 799032353, "item_id": 49939522, "url": "https://news.ycombinator.com/item?id=49939522", "by": "sleight42", "date": "2026-10-02", "stance": "compare", "quote": "either a blazing fast 8-bit Qwen 3.8 27b or a 4 or 5 bit Qwen 3.8 Flash Next", "excerpt": "I don't do that. I use my 2020 top end gaming PC with its 3090. I just ordered 128GB RAM for it. I'll use a single-chat/slot runtime like Strata or Freetoken. And then I'll be able to run either a blazing fast 8-bit Qwen 3.8 27b or a 4 or 5 bit Qwen 3.8 Flash Next. That's good enough for me for development. And then I'll have my current gaming rig with less RAM but better CPU and GPU run a reasonable smart tool agent", "context": "has a 2020 gaming PC with a 3090", "use_case": "local development model", "switched_from": "", "story": "Dots: Always-on agents"}, {"id": 799012386, "item_id": 49938274, "url": "https://news.ycombinator.com/item?id=49938274", "by": "vlowther", "date": "2026-10-02", "stance": "neutral", "quote": "when using Qwen 3.8 flash next", "excerpt": "It is pretty nifty. I spend some time over last weekend implementing fused TQ to allow for 1m context lengths on a 128 gb MacBook M5 Max when using Qwen 3.8 flash next (https://github.com/antirez/ds4/pull/1115 if you are interested). If I get bored I might port over the Metal kernels from oMLX -- the speed increase they have for the v0.7.0 release is amazeballs.", "context": "", "use_case": "long context local model", "switched_from": "", "story": "From the creator of Redis; run LLM locally with ds4"}, {"id": 799009536, "item_id": 49938096, "url": "https://news.ycombinator.com/item?id=49938096", "by": "roosterIllusi0n", "date": "2026-10-02", "stance": "recommend", "quote": "qwen3.8 is more than capable for software dev.", "excerpt": "I am using qwen3.8-27B-UD-Q3_K_XL on a 5080 16gb card with 16gb of system ram at around 43-45tok/s with 65k context. I am using that model to set up containers on a proxmox server with two b70s and 96gb ram. The Q3 model implemented 10 different chat models, 2 image models, and 7 web based harnesses so I can compare. It can rewrite any part to improve it. qwen3.8 is more than capable for software dev. The gemma model", "context": "runs it locally on a 5080 16GB", "use_case": "local software development", "switched_from": "", "story": "Several vulnerabilities have been discovered in the Linux kernel"}, {"id": 798997968, "item_id": 49937373, "url": "https://news.ycombinator.com/item?id=49937373", "by": "verdverm", "date": "2026-10-02", "stance": "recommend", "quote": "qwen3.8, kimi3, kimi2.7, GLM-5.3 are all good families I use in my coding team", "excerpt": "qwen3.8, kimi3, kimi2.7, GLM-5.3 are all good families I use in my coding team I'm mainly using flash varients, at least as the default, bump.up to stronger model as needed (less often these days)", "context": "", "use_case": "coding team", "switched_from": "", "story": "One month coding with GLM 5.3 Flash"}, {"id": 798966688, "item_id": 49935418, "url": "https://news.ycombinator.com/item?id=49935418", "by": "Roark66", "date": "2026-10-02", "stance": "recommend", "quote": "I've been running it locally since for real programming and I consider it pretty much the same as opus 4.6 in coding ability", "excerpt": "I wonder if August of this year will be remembered like that too. The very first \"opus like\" local AI model came out this August (Qwen 3.8 Flash Next). I've been running it locally since for real programming and I consider it pretty much the same as opus 4.6 in coding ability (it lacks a bit in the factual knowledge area). It even exceeds opus on some tasks. I've tried every hyped \"open source\" model before and all i", "context": "runs it on 6 RTX 3090s", "use_case": "local programming", "switched_from": "Claude Code", "story": "Several vulnerabilities have been discovered in the Linux kernel"}, {"id": 798939809, "item_id": 49933738, "url": "https://news.ycombinator.com/item?id=49933738", "by": "tingletech", "date": "2026-10-02", "stance": "recommend", "quote": "I use it with a local Qwen3.8 27B on llama-server", "excerpt": "I use it with a local Qwen3.8 27B on llama-server, but I don't know what it means for it exceed any window. you can use /slots on the llama-server if you want to get more up-to-date details on session token use. I always run models in their default context, which is 262144 for Qwen3.8 27B. I've run sessions in hax where it hit that cap multiple times and compressed the context down to 15% and continued with no proble", "context": "", "use_case": "local model for agent sessions", "switched_from": "", "story": "Pi 1.0"}, {"id": 798868402, "item_id": 49929275, "url": "https://news.ycombinator.com/item?id=49929275", "by": "pettijohn", "date": "2026-10-02", "stance": "recommend", "quote": "achieving flagship quality without the price tag (albeit slower)", "excerpt": "I use PI WEB (https://pi-web.dev, there's more than one with that name) to orchestrate remotely. Pi lets me run local models (currently Qwen 3.8 Flash Next on Strix Halo 128GB) and flagship models (with subscription auth, not API pricing) side by side. I typically ask GPT 6.1 Sol to review requirements and then spawn a subsession with Qwen, review Qwen's work, ask Qwen to fix. About 95% success with a single pass lik", "context": "runs local models on Strix Halo 128GB", "use_case": "local implementation model", "switched_from": "", "story": "Pi 1.0"}, {"id": 798837843, "item_id": 49927365, "url": "https://news.ycombinator.com/item?id=49927365", "by": "razster", "date": "2026-10-01", "stance": "recommend", "quote": "Local model is Qwen3.6 35B-A3B and Qwen3.8 27B UD.", "excerpt": "I use my Pi to organize my graphic design assets as well as update documentation for client projects. It also is my daily driver for a story I'm working on and thanks to a ton of great skills created it has become more powerful. Local model is Qwen3.6 35B-A3B and Qwen3.8 27B UD. I love the compressing function, keeps on pushing. I've tried Oh-my-pi but it's too heavy for me and eats up 6% of my context on start. Also", "context": "", "use_case": "local model for Pi", "switched_from": "", "story": "Pi 1.0"}, {"id": 797679840, "item_id": 49854990, "url": "https://news.ycombinator.com/item?id=49854990", "by": "praash", "date": "2026-09-26", "stance": "recommend", "quote": "Qwen 3.8 has been the first local model I've tried that has consistently given usable and impressive results.", "excerpt": "Qwen 3.8 27b-mlx running locally on Apple M4 48 GB. I only use local LLMs, mostly for getting technical suggestions and code snippet samples. Qwen 3.8 has been the first local model I've tried that has consistently given usable and impressive results.", "context": "only uses local LLMs on Apple M4 48 GB", "use_case": "local technical suggestions and code snippets", "switched_from": "", "story": "Ask HN: Which model do you use for work?"}, {"id": 794167632, "item_id": 49635477, "url": "https://news.ycombinator.com/item?id=49635477", "by": "anyfoo", "date": "2026-09-09", "stance": "warn", "quote": "Qwen3.8 27B does not seem to be good enough to implement the simulator with all its intricacies and tiny subtleties correctly", "excerpt": "You're probably partly right (though I use the same context size for all models, so there still is a difference in the model/quantization itself), but I definitely see it \"thinking itself into an extremely expensive and error-prone approach without backing out\". A common \"failure mode\" of my benchmark with