{"slug": "qwen-coder", "name": "Qwen Coder", "category": "ai", "category_name": "AI models and assistants", "description": "Alibaba's family of open-weight code-specialized large language models from the Qwen series.", "recommenders": 14, "commenters": 23, "warners": 2, "score": 12.375, "rising": -3.25, "first": "2025-11-01", "latest": "2026-10-22", "why": "HN commenters like it as a local, private coding model that runs on consumer hardware and works with agents like OpenCode.", "why_is_quote": false, "has_verdict": true, "sentiment": {"recommend": 14, "neutral": 7, "compare": 0, "warn": 2}, "spark": [0, 0, 1, 6, 1, 0, 2, 2, 1, 0, 0, 1], "listed": true, "loved": ["Local and private", "Good results"], "caveat": "Below frontier", "short_quote": "Qwen2.5-Coder 32B is genuinely impressive for local inference", "icon": true, "domain": "simonwillison.net", "github_repo": null, "official_url": "https://github.com/QwenLM/Qwen3-Coder", "pricing": "open source", "aliases": ["Qwen3-Coder-Next", "Qwen Coder", "Qwen2.5-Coder", "Qwen3 Coder", "Qwen3-Coder", "Qwen3 Coder Next", "Qwen Coder Next", "Qwen 2.5 Coder", "Qwen 3 Coder"], "verdict": [{"text": "HN commenters like it as a local, private coding model", "cites": [{"id": 771715793, "item_id": 48232237, "url": "https://news.ycombinator.com/item?id=48232237", "by": "alpha-pixel", "date": "2026-05-21", "stance": "recommend", "quote": "if you want something local and private Qwen 2.5 coder has been good at this time.", "excerpt": "For raw coding speed and handling large codebases claude 3.5 sonnet still feels unmatched. But if you want something local and private Qwen 2.5 coder has been good at this time.", "context": "", "use_case": "local private coding", "switched_from": "", "story": "Ask HN: What the Best AI for Coding?"}, {"id": 755287440, "item_id": 47205465, "url": "https://news.ycombinator.com/item?id=47205465", "by": "jedisct1", "date": "2026-03-01", "stance": "recommend", "quote": "Qwen3-Coder-Next also remains amazing as a local model.", "excerpt": "Qwen3-Coder-Next also remains amazing as a local model. If you want to use small models for coding, I'd highly recommend Swival https://swival.dev which was explicitly optimized for these.", "context": "", "use_case": "local coding model", "switched_from": "", "story": "Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers"}, {"id": 803204464, "item_id": 50200279, "url": "https://news.ycombinator.com/item?id=50200279", "by": "moooff", "date": "2026-10-22", "stance": "recommend", "quote": "it worked surprisingly well", "excerpt": "I was using it in an air gapped environment vs a local inference engine running a qwen-coder-next. Installtion of harnesses on the machine and other things were ruled out because of security constraints. There i used it to analyze an old visual basic application which creates custom pdf files. it worked surprisingly well. it took some rounds but in the end it got all requirements and scaffolded me a skeleton for a py", "context": "works in an air gapped environment", "use_case": "analyzing legacy code, rewrite scaffolding", "switched_from": "", "story": "OpenShell \u2013 A safe, private runtime for fleets of autonomous AI agents"}, {"id": 770348865, "item_id": 48146804, "url": "https://news.ycombinator.com/item?id=48146804", "by": "rsolva", "date": "2026-05-15", "stance": "recommend", "quote": "currently Qwen3.5 122B A10B and Qwen3 Coder Next", "excerpt": "In our company of 24 employees, we get by with two DGX Sparks. We don't use AI heavily, but each Spark can serve about 6-8 concurrent requests with a full context lenght of 256k, which is decent. We get about ~35 t/s depending on the model we use (currently Qwen3.5 122B A10B and Qwen3 Coder Next), but we might set up a smaller model too for simpler tasks. This works for us and will work for years to come. It is not S", "context": "company of 24 employees using DGX Sparks", "use_case": "local coding assistance", "switched_from": "", "story": "Cut Off: why access to frontier AI will soon be scarce and selective"}]}, {"text": " that runs on consumer hardware and works with agents like OpenCode.", "cites": [{"id": 746893040, "item_id": 46680815, "url": "https://news.ycombinator.com/item?id=46680815", "by": "dajonker", "date": "2026-01-19", "stance": "recommend", "quote": "So far Qwen3-coder gives the me best results.", "excerpt": "Great, I've been experimenting with OpenCode and running local 30B-A3B models on llama.cpp (4 bit) on a 32 GB GPU so there's plenty of VRAM left for 128k context. So far Qwen3-coder gives the me best results. Nemotron 3 Nano is supposed to benchmark better but it doesn't really show for the kind of work I throw at it, mostly \"write tests for this and that method which are not covered yet\". Will give this a try once s", "context": "runs local models on a 32 GB GPU", "use_case": "writing tests locally", "switched_from": "", "story": "GLM-4.7-Flash"}, {"id": 779498417, "item_id": 48718651, "url": "https://news.ycombinator.com/item?id=48718651", "by": "api", "date": "2026-06-29", "stance": "recommend", "quote": "This was with a 30b model at 4-bit quantization, so pretty tiny by LLM standards.", "excerpt": "Have you tried them recently? They absolutely 100% produce usable results for large classes of problems. Usually problems on the more \"boring\" side, but that's most code. Even local ones are getting \"usable.\" As an experiment I used OpenCode + Qwen coder to code a text based side scroller game in Rust. Left my laptop on and went to bed and in the morning I had a passable demo that looked like the old DOS game ZZT. Th", "context": "", "use_case": "local code generation", "switched_from": "", "story": "We need tech news sources which exclude AI"}, {"id": 776320385, "item_id": 48520024, "url": "https://news.ycombinator.com/item?id=48520024", "by": "bachmeier", "date": "2026-06-13", "stance": "recommend", "quote": "I run Gemma 4 26b (mixture of experts) and Qwen 3 coder with Ollama.", "excerpt": "> The upfront cost is steep and the models you can actually run at home are weaker than what the frontier labs ship, so this only pays off if you can keep the rig busy with long running tasks where a slower, cheaper model grinds away overnight. Most people can\u2019t keep a home machine that loaded, and the hardware you buy today may look like a bad bet in a year. Oh, so this is not a post about AI coding at home. It's ab", "context": "64 GB RAM, no GPU", "use_case": "local AI coding", "switched_from": "", "story": "AI Coding at Home Without Going Broke"}, {"id": 753243152, "item_id": 47077697, "url": "https://news.ycombinator.com/item?id=47077697", "by": "sosodev", "date": "2026-02-19", "stance": "recommend", "quote": "I think it might actually be the best option for agentic coding on the 395+ for now.", "excerpt": "Thanks for the additional info. I suspected that MiniMax M2.5 might be a bit too much for this board. 230B-A10B is just a lot to ask of the 395+ even with aggressive quantization. Particularly when you consider that the model is going to spend a lot of tokens thinking and that will eat into the comparatively smaller context window. I switched from the Unsloth 4-bit quant of Qwen3 Coder Next to the official 4-bit quan", "context": "running local models on a 395+ board", "use_case": "local agentic coding", "switched_from": "Unsloth 4-bit quant", "story": "Why AI Velocity Is Becoming a Debt Accelerator"}]}, {"text": " The most common caveat is that it trails cloud frontier models such as Claude and Codex,", "cites": [{"id": 787399441, "item_id": 49212465, "url": "https://news.ycombinator.com/item?id=49212465", "by": "joe-dakroub", "date": "2026-08-07", "stance": "neutral", "quote": "various local models like Qwen Coder Next to work on various tasks", "excerpt": "I use Bionic (https://lmstudio.ai) and various local models like Qwen Coder Next to work on various tasks. Honestly, none of the models that are runnable on a typical MacBook Pro come close to the cloud-based frontier models (open or closed) but Bionic provides a nice experience and makes trying new models out trivial.", "context": "", "use_case": "local coding tasks", "switched_from": "", "story": "Ask HN: How is everyone using Local LLMs?"}, {"id": 746893040, "item_id": 46680815, "url": "https://news.ycombinator.com/item?id=46680815", "by": "dajonker", "date": "2026-01-19", "stance": "recommend", "quote": "So far Qwen3-coder gives the me best results.", "excerpt": "Great, I've been experimenting with OpenCode and running local 30B-A3B models on llama.cpp (4 bit) on a 32 GB GPU so there's plenty of VRAM left for 128k context. So far Qwen3-coder gives the me best results. Nemotron 3 Nano is supposed to benchmark better but it doesn't really show for the kind of work I throw at it, mostly \"write tests for this and that method which are not covered yet\". Will give this a try once s", "context": "runs local models on a 32 GB GPU", "use_case": "writing tests locally", "switched_from": "", "story": "GLM-4.7-Flash"}, {"id": 771715793, "item_id": 48232237, "url": "https://news.ycombinator.com/item?id=48232237", "by": "alpha-pixel", "date": "2026-05-21", "stance": "recommend", "quote": "if you want something local and private Qwen 2.5 coder has been good at this time.", "excerpt": "For raw coding speed and handling large codebases claude 3.5 sonnet still feels unmatched. But if you want something local and private Qwen 2.5 coder has been good at this time.", "context": "", "use_case": "local private coding", "switched_from": "", "story": "Ask HN: What the Best AI for Coding?"}]}, {"text": " and some report bad code from small variants or unreliable tool calling.", "cites": [{"id": 802070225, "item_id": 50129389, "url": "https://news.ycombinator.com/item?id=50129389", "by": "testerius", "date": "2026-10-17", "stance": "warn", "quote": "However this model generates bad code", "excerpt": "It sounds great and I hope local LLM/AI will be more powerful and available even on old hardware. I tried llama.cpp with offloading \"all what could be\" to my GPU (GTX 750 Ti 2 GB). I could run local Qwen2.5-Coder quite fast (I do not remember but something like 25-30 t/s). However this model generates bad code (I tried Python and JavaScript). So, for my setup it is better to chat with free ChatGPT and do copying-past", "context": "", "use_case": "local coding", "switched_from": "", "story": "Show HN: Llmspeed \u2013 feel how fast a local LLM runs before you buy the hardware"}, {"id": 732508576, "item_id": 45781786, "url": "https://news.ycombinator.com/item?id=45781786", "by": "Balinares", "date": "2025-11-01", "stance": "neutral", "quote": "", "excerpt": "Does Qwen3 Coder do a good job invoking its tools as appropriate for you? Under continue.dev at least, I've found I need to remind it constantly.", "context": "", "use_case": "local coding assistant", "switched_from": "", "story": "Ask HN: Who uses open LLMs and coding assistants locally? Share setup and laptop"}]}], "praise": [{"label": "Local and private", "summary": "Commenters value running it locally, sometimes air-gapped, so code and data stay in-house.", "quotes": [{"id": 771715793, "item_id": 48232237, "url": "https://news.ycombinator.com/item?id=48232237", "by": "alpha-pixel", "date": "2026-05-21", "stance": "recommend", "quote": "if you want something local and private Qwen 2.5 coder has been good at this time.", "excerpt": "For raw coding speed and handling large codebases claude 3.5 sonnet still feels unmatched. But if you want something local and private Qwen 2.5 coder has been good at this time.", "context": "", "use_case": "local private coding", "switched_from": "", "story": "Ask HN: What the Best AI for Coding?"}, {"id": 803204464, "item_id": 50200279, "url": "https://news.ycombinator.com/item?id=50200279", "by": "moooff", "date": "2026-10-22", "stance": "recommend", "quote": "it worked surprisingly well", "excerpt": "I was using it in an air gapped environment vs a local inference engine running a qwen-coder-next. Installtion of harnesses on the machine and other things were ruled out because of security constraints. There i used it to analyze an old visual basic application which creates custom pdf files. it worked surprisingly well. it took some rounds but in the end it got all requirements and scaffolded me a skeleton for a py", "context": "works in an air gapped environment", "use_case": "analyzing legacy code, rewrite scaffolding", "switched_from": "", "story": "OpenShell \u2013 A safe, private runtime for fleets of autonomous AI agents"}, {"id": 770348865, "item_id": 48146804, "url": "https://news.ycombinator.com/item?id=48146804", "by": "rsolva", "date": "2026-05-15", "stance": "recommend", "quote": "currently Qwen3.5 122B A10B and Qwen3 Coder Next", "excerpt": "In our company of 24 employees, we get by with two DGX Sparks. We don't use AI heavily, but each Spark can serve about 6-8 concurrent requests with a full context lenght of 256k, which is decent. We get about ~35 t/s depending on the model we use (currently Qwen3.5 122B A10B and Qwen3 Coder Next), but we might set up a smaller model too for simpler tasks. This works for us and will work for years to come. It is not S", "context": "company of 24 employees using DGX Sparks", "use_case": "local coding assistance", "switched_from": "", "story": "Cut Off: why access to frontier AI will soon be scarce and selective"}]}, {"label": "Good results", "summary": "Users report strong coding output for a local model, including best results among similar small models.", "quotes": [{"id": 746893040, "item_id": 46680815, "url": "https://news.ycombinator.com/item?id=46680815", "by": "dajonker", "date": "2026-01-19", "stance": "recommend", "quote": "So far Qwen3-coder gives the me best results.", "excerpt": "Great, I've been experimenting with OpenCode and running local 30B-A3B models on llama.cpp (4 bit) on a 32 GB GPU so there's plenty of VRAM left for 128k context. So far Qwen3-coder gives the me best results. Nemotron 3 Nano is supposed to benchmark better but it doesn't really show for the kind of work I throw at it, mostly \"write tests for this and that method which are not covered yet\". Will give this a try once s", "context": "runs local models on a 32 GB GPU", "use_case": "writing tests locally", "switched_from": "", "story": "GLM-4.7-Flash"}, {"id": 755287440, "item_id": 47205465, "url": "https://news.ycombinator.com/item?id=47205465", "by": "jedisct1", "date": "2026-03-01", "stance": "recommend", "quote": "Qwen3-Coder-Next also remains amazing as a local model.", "excerpt": "Qwen3-Coder-Next also remains amazing as a local model. If you want to use small models for coding, I'd highly recommend Swival https://swival.dev which was explicitly optimized for these.", "context": "", "use_case": "local coding model", "switched_from": "", "story": "Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers"}, {"id": 779498417, "item_id": 48718651, "url": "https://news.ycombinator.com/item?id=48718651", "by": "api", "date": "2026-06-29", "stance": "recommend", "quote": "This was with a 30b model at 4-bit quantization, so pretty tiny by LLM standards.", "excerpt": "Have you tried them recently? They absolutely 100% produce usable results for large classes of problems. Usually problems on the more \"boring\" side, but that's most code. Even local ones are getting \"usable.\" As an experiment I used OpenCode + Qwen coder to code a text based side scroller game in Rust. Left my laptop on and went to bed and in the morning I had a passable demo that looked like the old DOS game ZZT. Th", "context": "", "use_case": "local code generation", "switched_from": "", "story": "We need tech news sources which exclude AI"}]}, {"label": "Consumer hardware", "summary": "It runs on laptops, single GPUs and even a 64 GB RAM machine without a GPU.", "quotes": [{"id": 776320385, "item_id": 48520024, "url": "https://news.ycombinator.com/item?id=48520024", "by": "bachmeier", "date": "2026-06-13", "stance": "recommend", "quote": "I run Gemma 4 26b (mixture of experts) and Qwen 3 coder with Ollama.", "excerpt": "> The upfront cost is steep and the models you can actually run at home are weaker than what the frontier labs ship, so this only pays off if you can keep the rig busy with long running tasks where a slower, cheaper model grinds away overnight. Most people can\u2019t keep a home machine that loaded, and the hardware you buy today may look like a bad bet in a year. Oh, so this is not a post about AI coding at home. It's ab", "context": "64 GB RAM, no GPU", "use_case": "local AI coding", "switched_from": "", "story": "AI Coding at Home Without Going Broke"}, {"id": 746893040, "item_id": 46680815, "url": "https://news.ycombinator.com/item?id=46680815", "by": "dajonker", "date": "2026-01-19", "stance": "recommend", "quote": "So far Qwen3-coder gives the me best results.", "excerpt": "Great, I've been experimenting with OpenCode and running local 30B-A3B models on llama.cpp (4 bit) on a 32 GB GPU so there's plenty of VRAM left for 128k context. So far Qwen3-coder gives the me best results. Nemotron 3 Nano is supposed to benchmark better but it doesn't really show for the kind of work I throw at it, mostly \"write tests for this and that method which are not covered yet\". Will give this a try once s", "context": "runs local models on a 32 GB GPU", "use_case": "writing tests locally", "switched_from": "", "story": "GLM-4.7-Flash"}, {"id": 779498417, "item_id": 48718651, "url": "https://news.ycombinator.com/item?id=48718651", "by": "api", "date": "2026-06-29", "stance": "recommend", "quote": "This was with a 30b model at 4-bit quantization, so pretty tiny by LLM standards.", "excerpt": "Have you tried them recently? They absolutely 100% produce usable results for large classes of problems. Usually problems on the more \"boring\" side, but that's most code. Even local ones are getting \"usable.\" As an experiment I used OpenCode + Qwen coder to code a text based side scroller game in Rust. Left my laptop on and went to bed and in the morning I had a passable demo that looked like the old DOS game ZZT. Th", "context": "", "use_case": "local code generation", "switched_from": "", "story": "We need tech news sources which exclude AI"}]}, {"label": "Agentic coding", "summary": "It is used with OpenCode and other agents for local agentic coding sessions.", "quotes": [{"id": 753243152, "item_id": 47077697, "url": "https://news.ycombinator.com/item?id=47077697", "by": "sosodev", "date": "2026-02-19", "stance": "recommend", "quote": "I think it might actually be the best option for agentic coding on the 395+ for now.", "excerpt": "Thanks for the additional info. I suspected that MiniMax M2.5 might be a bit too much for this board. 230B-A10B is just a lot to ask of the 395+ even with aggressive quantization. Particularly when you consider that the model is going to spend a lot of tokens thinking and that will eat into the comparatively smaller context window. I switched from the Unsloth 4-bit quant of Qwen3 Coder Next to the official 4-bit quan", "context": "running local models on a 395+ board", "use_case": "local agentic coding", "switched_from": "Unsloth 4-bit quant", "story": "Why AI Velocity Is Becoming a Debt Accelerator"}, {"id": 779498417, "item_id": 48718651, "url": "https://news.ycombinator.com/item?id=48718651", "by": "api", "date": "2026-06-29", "stance": "recommend", "quote": "This was with a 30b model at 4-bit quantization, so pretty tiny by LLM standards.", "excerpt": "Have you tried them recently? They absolutely 100% produce usable results for large classes of problems. Usually problems on the more \"boring\" side, but that's most code. Even local ones are getting \"usable.\" As an experiment I used OpenCode + Qwen coder to code a text based side scroller game in Rust. Left my laptop on and went to bed and in the morning I had a passable demo that looked like the old DOS game ZZT. Th", "context": "", "use_case": "local code generation", "switched_from": "", "story": "We need tech news sources which exclude AI"}, {"id": 780126098, "item_id": 48757881, "url": "https://news.ycombinator.com/item?id=48757881", "by": "Scrounger", "date": "2026-07-02", "stance": "neutral", "quote": "Qwen3-Coder-Next-GGUF:UD-Q4_K_XL to run local agentic coding sessions", "excerpt": "> I'm running these things in headless VMs What's your setup like and what do you use it for? I have a M2 Max MBP with plenty of ram and I use VSCode + Zoo Code plugin with Qwen3-Coder-Next-GGUF:UD-Q4_K_XL to run local agentic coding sessions, but I'm intrigued by being able to run headless as I could probably run multiple instances in parallel to do stuff? Like are you using UTM with some pre-built VM and a local LL", "context": "M2 Max MBP user", "use_case": "local agentic coding", "switched_from": "", "story": "ZCode: Claude Code from the Makers of GLM"}]}, {"label": "Speed", "summary": "Commenters describe decent speed, with prompts often returning in under 30 seconds.", "quotes": [{"id": 751305376, "item_id": 46956586, "url": "https://news.ycombinator.com/item?id=46956586", "by": "0xbadcafebee", "date": "2026-02-09", "stance": "recommend", "quote": "Qwen3 coder is pretty good and decent speed.", "excerpt": "Local models are decent now. Qwen3 coder is pretty good and decent speed. I use smaller models (qwen2.5:1.5b) with keyboard shortcuts and speech to text to ask for man page entries, and get 'em back faster than my internet connection and a \"robust\" frontier model does. And web search/RAG hides a multitude of sins. \"Using anything other than the frontier models is actively harmful\" - so how come I'm getting solid resu", "context": "", "use_case": "local coding model", "switched_from": "", "story": "Eight More Months of Agents"}, {"id": 784331136, "item_id": 49020696, "url": "https://news.ycombinator.com/item?id=49020696", "by": "gbuk2013", "date": "2026-07-23", "stance": "recommend", "quote": "I run Qwen3-Coder-Next-UD-Q4_K_XL and other than the initial wait to initialise context", "excerpt": "I run Qwen3-Coder-Next-UD-Q4_K_XL and other than the initial wait to initialise context (which takes less than 2 minutes) subsequent prompts return in less than a minute, usually less than 30s. If your performance is significantly slower then you are probably doing it in CPU - there was some fiddling required to get it to use GPU (I use llama.cpp)", "context": "", "use_case": "local coding inference", "switched_from": "", "story": "New Framework Desktop Option with AMD Ryzen AI Max+ Pro 495 and 192GB Memory"}, {"id": 750208240, "item_id": 46888015, "url": "https://news.ycombinator.com/item?id=46888015", "by": "mark_l_watson", "date": "2026-02-04", "stance": "recommend", "quote": "A 3B resident parameter MOE allows absolutely huge savings on inference costs.", "excerpt": "A 3B resident parameter MOE allows absolutely huge savings on inference costs. I use a cloud provider for models to large to run locally, can\u2019t wit for them to support qwen3-coder-next hopefully in a few days. So much expensive inference is provided free or at large discounts - that craziness should end.", "context": "", "use_case": "coding model inference", "switched_from": "", "story": "Qwen3-Coder-Next"}]}], "caveats": [{"label": "Below frontier", "summary": "Local runs do not match cloud frontier models like Claude Sonnet or Codex.", "quotes": [{"id": 787399441, "item_id": 49212465, "url": "https://news.ycombinator.com/item?id=49212465", "by": "joe-dakroub", "date": "2026-08-07", "stance": "neutral", "quote": "various local models like Qwen Coder Next to work on various tasks", "excerpt": "I use Bionic (https://lmstudio.ai) and various local models like Qwen Coder Next to work on various tasks. Honestly, none of the models that are runnable on a typical MacBook Pro come close to the cloud-based frontier models (open or closed) but Bionic provides a nice experience and makes trying new models out trivial.", "context": "", "use_case": "local coding tasks", "switched_from": "", "story": "Ask HN: How is everyone using Local LLMs?"}, {"id": 746893040, "item_id": 46680815, "url": "https://news.ycombinator.com/item?id=46680815", "by": "dajonker", "date": "2026-01-19", "stance": "recommend", "quote": "So far Qwen3-coder gives the me best results.", "excerpt": "Great, I've been experimenting with OpenCode and running local 30B-A3B models on llama.cpp (4 bit) on a 32 GB GPU so there's plenty of VRAM left for 128k context. So far Qwen3-coder gives the me best results. Nemotron 3 Nano is supposed to benchmark better but it doesn't really show for the kind of work I throw at it, mostly \"write tests for this and that method which are not covered yet\". Will give this a try once s", "context": "runs local models on a 32 GB GPU", "use_case": "writing tests locally", "switched_from": "", "story": "GLM-4.7-Flash"}, {"id": 771715793, "item_id": 48232237, "url": "https://news.ycombinator.com/item?id=48232237", "by": "alpha-pixel", "date": "2026-05-21", "stance": "recommend", "quote": "if you want something local and private Qwen 2.5 coder has been good at this time.", "excerpt": "For raw coding speed and handling large codebases claude 3.5 sonnet still feels unmatched. But if you want something local and private Qwen 2.5 coder has been good at this time.", "context": "", "use_case": "local private coding", "switched_from": "", "story": "Ask HN: What the Best AI for Coding?"}]}, {"label": "Small-model quality", "summary": "One user found Qwen2.5-Coder on a 2 GB GPU generated bad Python and JavaScript.", "quotes": [{"id": 802070225, "item_id": 50129389, "url": "https://news.ycombinator.com/item?id=50129389", "by": "testerius", "date": "2026-10-17", "stance": "warn", "quote": "However this model generates bad code", "excerpt": "It sounds great and I hope local LLM/AI will be more powerful and available even on old hardware. I tried llama.cpp with offloading \"all what could be\" to my GPU (GTX 750 Ti 2 GB). I could run local Qwen2.5-Coder quite fast (I do not remember but something like 25-30 t/s). However this model generates bad code (I tried Python and JavaScript). So, for my setup it is better to chat with free ChatGPT and do copying-past", "context": "", "use_case": "local coding", "switched_from": "", "story": "Show HN: Llmspeed \u2013 feel how fast a local LLM runs before you buy the hardware"}]}, {"label": "Tool calling", "summary": "One user had to remind it constantly to invoke its tools under continue.dev.", "quotes": [{"id": 732508576, "item_id": 45781786, "url": "https://news.ycombinator.com/item?id=45781786", "by": "Balinares", "date": "2025-11-01", "stance": "neutral", "quote": "", "excerpt": "Does Qwen3 Coder do a good job invoking its tools as appropriate for you? Under continue.dev at least, I've found I need to remind it constantly.", "context": "", "use_case": "local coding assistant", "switched_from": "", "story": "Ask HN: Who uses open LLMs and coding assistants locally? Share setup and laptop"}]}, {"label": "Setup tuning", "summary": "Getting GPU use or higher speeds required fiddling with inference settings.", "quotes": [{"id": 784331136, "item_id": 49020696, "url": "https://news.ycombinator.com/item?id=49020696", "by": "gbuk2013", "date": "2026-07-23", "stance": "recommend", "quote": "I run Qwen3-Coder-Next-UD-Q4_K_XL and other than the initial wait to initialise context", "excerpt": "I run Qwen3-Coder-Next-UD-Q4_K_XL and other than the initial wait to initialise context (which takes less than 2 minutes) subsequent prompts return in less than a minute, usually less than 30s. If your performance is significantly slower then you are probably doing it in CPU - there was some fiddling required to get it to use GPU (I use llama.cpp)", "context": "", "use_case": "local coding inference", "switched_from": "", "story": "New Framework Desktop Option with AMD Ryzen AI Max+ Pro 495 and 192GB Memory"}, {"id": 776141219, "item_id": 48508826, "url": "https://news.ycombinator.com/item?id=48508826", "by": "LoganDark", "date": "2026-06-12", "stance": "neutral", "quote": "custom Burn inference for Qwen3-Coder-Next", "excerpt": "I poured a couple days into custom Burn inference for Qwen3-Coder-Next only to find it doesn't come with a speculative decoder, so on my M4 Max I can't push it much further than 120t/s. That's still kinda slow, though still faster than llama.cpp's 70.9t/s and MLX's 80.6t/s with the same model. Claude Fable 5 is recommending I use the Qwen3 MTP -- I worry that will compromise the quality somewhat, but might give it a ", "context": "", "use_case": "local coding model", "switched_from": "", "story": "How to Setup a Local Coding Agent on macOS"}]}, {"label": "Hosted rate limits", "summary": "On a hosted subscription, tools quickly hit per-minute request limits with the 480B model.", "quotes": [{"id": 733714084, "item_id": 45857130, "url": "https://news.ycombinator.com/item?id=45857130", "by": "sheepscreek", "date": "2025-11-08", "stance": "warn", "quote": "none of the \u201csupported\u201d tools I tried were able to use it in a way that didn\u2019t hit the per minute request limit in a few seconds.", "excerpt": "I used their $50 plan and with the previously offered Qwen3 coder 480B. While fast - none of the \u201csupported\u201d tools I tried were able to use it in a way that didn\u2019t hit the per minute request limit in a few seconds. It was incredibly frustrating. For the record, I tried OpenCoder, VSCode, Quen Coder CLI, octofriend and a few others I don\u2019t remember. Fast forward to now, when GLM 4.6 has replaced Qwen3 coder in their s", "context": "", "use_case": "coding via Cerebras", "switched_from": "", "story": "Cerebras Code now supports GLM 4.6 at 1000 tokens/sec"}]}], "who": [{"by": "moooff", "context": "works in an air gapped environment", "url": "https://news.ycombinator.com/item?id=50200279", "date": "2026-10-22"}, {"by": "Scrounger", "context": "M2 Max MBP user", "url": "https://news.ycombinator.com/item?id=48757881", "date": "2026-07-02"}, {"by": "bachmeier", "context": "64 GB RAM, no GPU", "url": "https://news.ycombinator.com/item?id=48520024", "date": "2026-06-13"}, {"by": "rsolva", "context": "company of 24 employees using DGX Sparks", "url": "https://news.ycombinator.com/item?id=48146804", "date": "2026-05-15"}, {"by": "sosodev", "context": "running local models on a 395+ board", "url": "https://news.ycombinator.com/item?id=47077697", "date": "2026-02-19"}, {"by": "josefcub", "context": "runs a 256GB M3 Ultra Mac Studio", "url": "https://news.ycombinator.com/item?id=46915397", "date": "2026-02-06"}, {"by": "dajonker", "context": "runs local models on a 32 GB GPU", "url": "https://news.ycombinator.com/item?id=46680815", "date": "2026-01-19"}], "tags": ["for local inference", "for air-gapped environments", "for Apple Silicon Macs", "for small teams", "for code completion"], "use_cases": ["local coding model", "local agentic coding", "local code generation", "local coding assistant", "analyzing legacy code, rewrite scaffolding", "local coding"], "compared": [{"slug": "llama-cpp", "name": "llama.cpp", "co": 5, "switched_from": 0, "switched_to": 0}, {"slug": "glm", "name": "GLM", "co": 3, "switched_from": 0, "switched_to": 1}, {"slug": "claude", "name": "Claude", "co": 3, "switched_from": 0, "switched_to": 0}, {"slug": "opencode", "name": "OpenCode", "co": 3, "switched_from": 0, "switched_to": 0}, {"slug": "codex", "name": "Codex", "co": 3, "switched_from": 0, "switched_to": 0}, {"slug": "github-copilot", "name": "GitHub Copilot", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "chatgpt", "name": "ChatGPT", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "visual-studio-code", "name": "Visual Studio Code", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=45771870", "title": "Ask HN: Who uses open LLMs and coding assistants locally? Share setup and laptop", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=50198675", "title": "OpenShell \u2013 A safe, private runtime for fleets of autonomous AI agents", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=50127979", "title": "Show HN: Llmspeed \u2013 feel how fast a local LLM runs before you buy the hardware", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49211806", "title": "Ask HN: How is everyone using Local LLMs?", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49019694", "title": "New Framework Desktop Option with AMD Ryzen AI Max+ Pro 495 and 192GB Memory", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48751752", "title": "ZCode: Claude Code from the Makers of GLM", "mentions": 1}], "sentiment_by_month": {"2025-11": {"neutral": 2, "warn": 1}, "2025-12": {}, "2026-01": {"neutral": 2, "recommend": 1}, "2026-02": {"recommend": 6}, "2026-03": {"recommend": 1}, "2026-04": {}, "2026-05": {"recommend": 2}, "2026-06": {"neutral": 1, "recommend": 2}, "2026-07": {"neutral": 1, "recommend": 1}, "2026-08": {"neutral": 1}, "2026-09": {}, "2026-10": {"warn": 1, "recommend": 1}}, "quotes": [{"id": 803204464, "item_id": 50200279, "url": "https://news.ycombinator.com/item?id=50200279", "by": "moooff", "date": "2026-10-22", "stance": "recommend", "quote": "it worked surprisingly well", "excerpt": "I was using it in an air gapped environment vs a local inference engine running a qwen-coder-next. Installtion of harnesses on the machine and other things were ruled out because of security constraints. There i used it to analyze an old visual basic application which creates custom pdf files. it worked surprisingly well. it took some rounds but in the end it got all requirements and scaffolded me a skeleton for a py", "context": "works in an air gapped environment", "use_case": "analyzing legacy code, rewrite scaffolding", "switched_from": "", "story": "OpenShell \u2013 A safe, private runtime for fleets of autonomous AI agents"}, {"id": 802070225, "item_id": 50129389, "url": "https://news.ycombinator.com/item?id=50129389", "by": "testerius", "date": "2026-10-17", "stance": "warn", "quote": "However this model generates bad code", "excerpt": "It sounds great and I hope local LLM/AI will be more powerful and available even on old hardware. I tried llama.cpp with offloading \"all what could be\" to my GPU (GTX 750 Ti 2 GB). I could run local Qwen2.5-Coder quite fast (I do not remember but something like 25-30 t/s). However this model generates bad code (I tried Python and JavaScript). So, for my setup it is better to chat with free ChatGPT and do copying-past", "context": "", "use_case": "local coding", "switched_from": "", "story": "Show HN: Llmspeed \u2013 feel how fast a local LLM runs before you buy the hardware"}, {"id": 787399441, "item_id": 49212465, "url": "https://news.ycombinator.com/item?id=49212465", "by": "joe-dakroub", "date": "2026-08-07", "stance": "neutral", "quote": "various local models like Qwen Coder Next to work on various tasks", "excerpt": "I use Bionic (https://lmstudio.ai) and various local models like Qwen Coder Next to work on various tasks. Honestly, none of the models that are runnable on a typical MacBook Pro come close to the cloud-based frontier models (open or closed) but Bionic provides a nice experience and makes trying new models out trivial.", "context": "", "use_case": "local coding tasks", "switched_from": "", "story": "Ask HN: How is everyone using Local LLMs?"}, {"id": 784331136, "item_id": 49020696, "url": "https://news.ycombinator.com/item?id=49020696", "by": "gbuk2013", "date": "2026-07-23", "stance": "recommend", "quote": "I run Qwen3-Coder-Next-UD-Q4_K_XL and other than the initial wait to initialise context", "excerpt": "I run Qwen3-Coder-Next-UD-Q4_K_XL and other than the initial wait to initialise context (which takes less than 2 minutes) subsequent prompts return in less than a minute, usually less than 30s. If your performance is significantly slower then you are probably doing it in CPU - there was some fiddling required to get it to use GPU (I use llama.cpp)", "context": "", "use_case": "local coding inference", "switched_from": "", "story": "New Framework Desktop Option with AMD Ryzen AI Max+ Pro 495 and 192GB Memory"}, {"id": 780126098, "item_id": 48757881, "url": "https://news.ycombinator.com/item?id=48757881", "by": "Scrounger", "date": "2026-07-02", "stance": "neutral", "quote": "Qwen3-Coder-Next-GGUF:UD-Q4_K_XL to run local agentic coding sessions", "excerpt": "> I'm running these things in headless VMs What's your setup like and what do you use it for? I have a M2 Max MBP with plenty of ram and I use VSCode + Zoo Code plugin with Qwen3-Coder-Next-GGUF:UD-Q4_K_XL to run local agentic coding sessions, but I'm intrigued by being able to run headless as I could probably run multiple instances in parallel to do stuff? Like are you using UTM with some pre-built VM and a local LL", "context": "M2 Max MBP user", "use_case": "local agentic coding", "switched_from": "", "story": "ZCode: Claude Code from the Makers of GLM"}, {"id": 779498417, "item_id": 48718651, "url": "https://news.ycombinator.com/item?id=48718651", "by": "api", "date": "2026-06-29", "stance": "recommend", "quote": "This was with a 30b model at 4-bit quantization, so pretty tiny by LLM standards.", "excerpt": "Have you tried them recently? They absolutely 100% produce usable results for large classes of problems. Usually problems on the more \"boring\" side, but that's most code. Even local ones are getting \"usable.\" As an experiment I used OpenCode + Qwen coder to code a text based side scroller game in Rust. Left my laptop on and went to bed and in the morning I had a passable demo that looked like the old DOS game ZZT. Th", "context": "", "use_case": "local code generation", "switched_from": "", "story": "We need tech news sources which exclude AI"}, {"id": 776320385, "item_id": 48520024, "url": "https://news.ycombinator.com/item?id=48520024", "by": "bachmeier", "date": "2026-06-13", "stance": "recommend", "quote": "I run Gemma 4 26b (mixture of experts) and Qwen 3 coder with Ollama.", "excerpt": "> The upfront cost is steep and the models you can actually run at home are weaker than what the frontier labs ship, so this only pays off if you can keep the rig busy with long running tasks where a slower, cheaper model grinds away overnight. Most people can\u2019t keep a home machine that loaded, and the hardware you buy today may look like a bad bet in a year. Oh, so this is not a post about AI coding at home. It's ab", "context": "64 GB RAM, no GPU", "use_case": "local AI coding", "switched_from": "", "story": "AI Coding at Home Without Going Broke"}, {"id": 776141219, "item_id": 48508826, "url": "https://news.ycombinator.com/item?id=48508826", "by": "LoganDark", "date": "2026-06-12", "stance": "neutral", "quote": "custom Burn inference for Qwen3-Coder-Next", "excerpt": "I poured a couple days into custom Burn inference for Qwen3-Coder-Next only to find it doesn't come with a speculative decoder, so on my M4 Max I can't push it much further than 120t/s. That's still kinda slow, though still faster than llama.cpp's 70.9t/s and MLX's 80.6t/s with the same model. Claude Fable 5 is recommending I use the Qwen3 MTP -- I worry that will compromise the quality somewhat, but might give it a ", "context": "", "use_case": "local coding model", "switched_from": "", "story": "How to Setup a Local Coding Agent on macOS"}, {"id": 771715793, "item_id": 48232237, "url": "https://news.ycombinator.com/item?id=48232237", "by": "alpha-pixel", "date": "2026-05-21", "stance": "recommend", "quote": "if you want something local and private Qwen 2.5 coder has been good at this time.", "excerpt": "For raw coding speed and handling large codebases claude 3.5 sonnet still feels unmatched. But if you want something local and private Qwen 2.5 coder has been good at this time.", "context": "", "use_case": "local private coding", "switched_from": "", "story": "Ask HN: What the Best AI for Coding?"}, {"id": 770348865, "item_id": 48146804, "url": "https://news.ycombinator.com/item?id=48146804", "by": "rsolva", "date": "2026-05-15", "stance": "recommend", "quote": "currently Qwen3.5 122B A10B and Qwen3 Coder Next", "excerpt": "In our company of 24 employees, we get by with two DGX Sparks. We don't use AI heavily, but each Spark can serve about 6-8 concurrent requests with a full context lenght of 256k, which is decent. We get about ~35 t/s depending on the model we use (currently Qwen3.5 122B A10B and Qwen3 Coder Next), but we might set up a smaller model too for simpler tasks. This works for us and will work for years to come. It is not S", "context": "company of 24 employees using DGX Sparks", "use_case": "local coding assistance", "switched_from": "", "story": "Cut Off: why access to frontier AI will soon be scarce and selective"}, {"id": 755287440, "item_id": 47205465, "url": "https://news.ycombinator.com/item?id=47205465", "by": "jedisct1", "date": "2026-03-01", "stance": "recommend", "quote": "Qwen3-Coder-Next also remains amazing as a local model.", "excerpt": "Qwen3-Coder-Next also remains amazing as a local model. If you want to use small models for coding, I'd highly recommend Swival https://swival.dev which was explicitly optimized for these.", "context": "", "use_case": "local coding model", "switched_from": "", "story": "Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers"}, {"id": 753379860, "item_id": 47086241, "url": "https://news.ycombinator.com/item?id=47086241", "by": "laramie_co", "date": "2026-02-20", "stance": "recommend", "quote": "Qwen2.5-Coder 32B is genuinely impressive for local inference", "excerpt": "Hey HN \u2014 maker here. Quick context on why I built this: I was personally spending ~$1,200/month on Claude API (I use it for everything \u2014 coding, writing, analysis). That\u2019s $14,400/year. Even if I dropped to a single $20/month sub, every query I send tells Anthropic what I\u2019m working on, what I\u2019m thinking about, what problems I have. I wanted Claude-quality inference without the cloud dependency. That\u2019s not possible ye", "context": "", "use_case": "local coding assistant", "switched_from": "", "story": "Show HN: VaultAI \u2013 42 AI models on a portable SSD, works offline ($399)"}, {"id": 753243152, "item_id": 47077697, "url": "https://news.ycombinator.com/item?id=47077697", "by": "sosodev", "date": "2026-02-19", "stance": "recommend", "quote": "I think it might actually be the best option for agentic coding on the 395+ for now.", "excerpt": "Thanks for the additional info. I suspected that MiniMax M2.5 might be a bit too much for this board. 230B-A10B is just a lot to ask of the 395+ even with aggressive quantization. Particularly when you consider that the model is going to spend a lot of tokens thinking and that will eat into the comparatively smaller context window. I switched from the Unsloth 4-bit quant of Qwen3 Coder Next to the official 4-bit quan", "context": "running local models on a 395+ board", "use_case": "local agentic coding", "switched_from": "Unsloth 4-bit quant", "story": "Why AI Velocity Is Becoming a Debt Accelerator"}, {"id": 752558464, "item_id": 47034904, "url": "https://news.ycombinator.com/item?id=47034904", "by": "andsoitis", "date": "2026-02-16", "stance": "recommend", "quote": "You can run smaller models locally that are pretty good at code gen.", "excerpt": "You can run smaller models locally that are pretty good at code gen. Give it a try. To get started: https://simonwillison.net/2024/Nov/12/qwen25-coder/ or https://simonwillison.net/2024/Dec/9/llama-33-70b/", "context": "", "use_case": "local code generation", "switched_from": "", "story": "Ask HN: What happens after the AI bubble bursts?"}, {"id": 751305376, "item_id": 46956586, "url": "https://news.ycombinator.com/item?id=46956586", "by": "0xbadcafebee", "date": "2026-02-09", "stance": "recommend", "quote": "Qwen3 coder is pretty good and decent speed.", "excerpt": "Local models are decent now. Qwen3 coder is pretty good and decent speed. I use smaller models (qwen2.5:1.5b) with keyboard shortcuts and speech to text to ask for man page entries, and get 'em back faster than my internet connection and a \"robust\" frontier model does. And web search/RAG hides a multitude of sins. \"Using anything other than the frontier models is actively harmful\" - so how come I'm getting solid resu", "context": "", "use_case": "local coding model", "switched_from": "", "story": "Eight More Months of Agents"}, {"id": 750646353, "item_id": 46915397, "url": "https://news.ycombinator.com/item?id=46915397", "by": "josefcub", "date": "2026-02-06", "stance": "recommend", "quote": "The smaller Qwen3 models (like qwen3-coder) I use in tandem", "excerpt": "I am! I moved from a shoebox Linux workstation with 32MB of RAM and a 12GB RTX 3060 to a 256GB M3 Ultra, mainly for unified memory. I've only had it a couple of months, but so far it's proving its worth in the quality of LLM output, even quantized. I generally run Qwen3-vl at 235b, at a Q4_K_M quantization level so that it fits, and it leaves me plenty of RAM for workstation tasks while delivering tokens at around 30", "context": "runs a 256GB M3 Ultra Mac Studio", "use_case": "local coding model", "switched_from": "", "story": "Ask HN: Anyone Using a Mac Studio for Local AI/LLM?"}, {"id": 750208240, "item_id": 46888015, "url": "https://news.ycombinator.com/item?id=46888015", "by": "mark_l_watson", "date": "2026-02-04", "stance": "recommend", "quote": "A 3B resident parameter MOE allows absolutely huge savings on inference costs.", "excerpt": "A 3B resident parameter MOE allows absolutely huge savings on inference costs. I use a cloud provider for models to large to run locally, can\u2019t wit for them to support qwen3-coder-next hopefully in a few days. So much expensive inference is provided free or at large discounts - that craziness should end.", "context": "", "use_case": "coding model inference", "switched_from": "", "story": "Qwen3-Coder-Next"}, {"id": 746893040, "item_id": 46680815, "url": "https://news.ycombinator.com/item?id=46680815", "by": "dajonker", "date": "2026-01-19", "stance": "recommend", "quote": "So far Qwen3-coder gives the me best results.", "excerpt": "Great, I've been experimenting with OpenCode and running local 30B-A3B models on llama.cpp (4 bit) on a 32 GB GPU so there's plenty of VRAM left for 128k context. So far Qwen3-coder gives the me best results. Nemotron 3 Nano is supposed to benchmark better but it doesn't really show for the kind of work I throw at it, mostly \"write tests for this and that method which are not covered yet\". Will give this a try once s", "context": "runs local models on a 32 GB GPU", "use_case": "writing tests locally", "switched_from": "", "story": "GLM-4.7-Flash"}, {"id": 744973635, "item_id": 46560852, "url": "https://news.ycombinator.com/item?id=46560852", "by": "delaminator", "date": "2026-01-09", "stance": "neutral", "quote": "I've got a Linux chatbot helper on the \"test this\" pile with Qwen Coder - not evaluated it", "excerpt": "I have a 3090 24Gb Twin Xenon 64Gb RAM sat on a machine in our server room. I do local AI with Qwen, Whisper and another I can't remember right now. These are all QWEN: We do AI Invoice OCR - PDF -> Image -> Excel. Works much better than other solutions because it has invoice context so it looks for particular data to extract and ignores others. Why local? I proved it worked, no need to send our data outside for proc", "context": "", "use_case": "Linux command helper", "switched_from": "", "story": "Ask HN: Who's running local AI workstations in 2026?"}, {"id": 744150742, "item_id": 46509421, "url": "https://news.ycombinator.com/item?id=46509421", "by": "Abishek_Muthian", "date": "2026-01-05", "stance": "neutral", "quote": "a custom agent for Qwen Coder via Llama.cpp [1]", "excerpt": "I just have a minimalist setup. Configuration - In external agents - I have Calude Code, Codex CLI, Gemini CLI and a custom agent for Qwen Coder via Llama.cpp [1] In MCP - I have fetch, brave-search, puppeteer In LLM Providers - I have configured Llama.cpp, LMStudio and OpenAI (Zed Agent can access models from any of these providers). Workflow - When I need LLM assist I mostly use just Claude Code for specific tasks,", "context": "", "use_case": "local coding model", "switched_from": "", "story": "I switched from VSCode to Zed"}, {"id": 733714084, "item_id": 45857130, "url": "https://news.ycombinator.com/item?id=45857130", "by": "sheepscreek", "date": "2025-11-08", "stance": "warn", "quote": "none of the \u201csupported\u201d tools I tried were able to use it in a way that didn\u2019t hit the per minute request limit in a few seconds.", "excerpt": "I used their $50 plan and with the previously offered Qwen3 coder 480B. While fast - none of the \u201csupported\u201d tools I tried were able to use it in a way that didn\u2019t hit the per minute request limit in a few seconds. It was incredibly frustrating. For the record, I tried OpenCoder, VSCode, Quen Coder CLI, octofriend and a few others I don\u2019t remember. Fast forward to now, when GLM 4.6 has replaced Qwen3 coder in their s", "context": "", "use_case": "coding via Cerebras", "switched_from": "", "story": "Cerebras Code now supports GLM 4.6 at 1000 tokens/sec"}, {"id": 732537860, "item_id": 45783616, "url": "https://news.ycombinator.com/item?id=45783616", "by": "toogle", "date": "2025-11-01", "stance": "neutral", "quote": "mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit", "excerpt": "I built a custom code-completion server: <https://github.com/toogle/mlx-dev-server>. The key advantage is that it cancels generation when you continue typing, so invalidated completions don\u2019t waste time. This makes completion latency predictable (about 1.5 seconds for me). My setup: - MacBook Pro (M3 Max) - Neovim - https://github.com/huggingface/llm.nvim Models I typically use: - mlx-community/DeepSeek-Coder-V2-Lite", "context": "", "use_case": "local code completion", "switched_from": "", "story": "Ask HN: Who uses open LLMs and coding assistants locally? Share setup and laptop"}, {"id": 732508576, "item_id": 45781786, "url": "https://news.ycombinator.com/item?id=45781786", "by": "Balinares", "date": "2025-11-01", "stance": "neutral", "quote": "", "excerpt": "Does Qwen3 Coder do a good job invoking its tools as appropriate for you? Under continue.dev at least, I've found I need to remind it constantly.", "context": "", "use_case": "local coding assistant", "switched_from": "", "story": "Ask HN: Who uses open LLMs and coding assistants locally? Share setup and laptop"}]}