{"slug": "ds4", "name": "ds4", "category": "ai", "category_name": "AI models and assistants", "description": "Local LLM inference engine by antirez for DeepSeek V4 and other large open models", "recommenders": 14, "commenters": 20, "warners": 0, "score": 13.997, "rising": 3.05, "first": "2026-06-12", "latest": "2026-10-17", "why": "HN commenters like it for running DeepSeek V4 Flash and other large open models locally on 128GB Macs, Strix Halo and DGX Spark machines, and several say it is solid enough for daily coding and agent work.", "why_is_quote": false, "has_verdict": true, "sentiment": {"recommend": 17, "neutral": 3, "compare": 3, "warn": 0}, "spark": [0, 0, 0, 0, 0, 0, 0, 8, 2, 0, 3, 4], "listed": true, "loved": ["Large local models", "Daily reliability"], "caveat": "Generation speed", "short_quote": "I use it as my main coding agent - so its running DS4 server on my 128gb mbp", "icon": true, "domain": null, "github_repo": "github.com/antirez/ds4", "official_url": "https://github.com/antirez/ds4", "pricing": "open source", "aliases": ["ds4", "DS4", "DwarfStar", "dwarfstar", "Dwarfstar"], "verdict": [{"text": "HN commenters like it for running DeepSeek V4 Flash and other large open models locally on 128GB Macs, Strix Halo and DGX Spark machines,", "cites": [{"id": 779590305, "item_id": 48724394, "url": "https://news.ycombinator.com/item?id=48724394", "by": "ljosifov", "date": "2026-06-29", "stance": "recommend", "quote": "DeepSeek-V4-Flash will fit, served by DwarfStar", "excerpt": "Running 27B dense model on M5 128GB is ok, but one can do better. On M5 128GB one can make use of the ram and use sparse MoE. For example, DeepSeek-V4-Flash will fit, served by DwarfStar (https://github.com/antirez/ds4). One will probably improve 2x the token/sec speed, given DS4F 13B activated params in the MoE are ~1/2 of the ~27B of the dense Qwen. 27B Of the Qwen fit even on a cheaper 24GB card, e.g. amd 7900xtx ", "context": "", "use_case": "serving DeepSeek V4 Flash locally", "switched_from": "", "story": "Qwen 3.6 27B is the sweet spot for local development"}, {"id": 776173857, "item_id": 48510866, "url": "https://news.ycombinator.com/item?id=48510866", "by": "jumploops", "date": "2026-06-12", "stance": "recommend", "quote": "running via antirez's ds4", "excerpt": "I've been quite impressed with DeepSeek v4 Flash running via antirez's ds4[0]. It feels like a GPT-4 class model in terms of \"stored knowledge\" but is better at long-horizon tool calling than any of the GPT-4 class models. Running on a 128GB MBP M4 Max, I'm getting ~24 t/s on generation and ~200 t/s on prefill. I was expecting it to feel slow, and it certainly does when e.g. generating code, but it's surprisingly use", "context": "128GB MBP M4 Max", "use_case": "running DeepSeek V4 Flash locally", "switched_from": "", "story": "How to Setup a Local Coding Agent on macOS"}, {"id": 779260672, "item_id": 48703792, "url": "https://news.ycombinator.com/item?id=48703792", "by": "pixelpoet", "date": "2026-06-27", "stance": "recommend", "quote": "have been extremely excited about Antirez's (Redis author) work on DS4", "excerpt": "I have two 128gb Strix Halos and have been extremely excited about Antirez's (Redis author) work on DS4, especially with 4bit quant using two machines: https://github.com/antirez/ds4 Right now the speed isn't good for GLM 5.2, Deepseek V4 Flash speed is okay for me (actually reading the output) and quite usable. See kyuz0's great recent video here: https://www.youtube.com/watch?v=PkKXm_mKCCM With a bit more speed and", "context": "owns two Strix Halo machines", "use_case": "local LLM inference on Strix Halo", "switched_from": "", "story": "AMD Strix Halo RDMA Cluster Setup Guide"}, {"id": 777910720, "item_id": 48619420, "url": "https://news.ycombinator.com/item?id=48619420", "by": "wolttam", "date": "2026-06-21", "stance": "recommend", "quote": "I started with antirez' DwarfStar[1] on one spark and that (~11-14tok/s generation, ~300-400 tok/s prompt processing) was enough of a taste for me", "excerpt": "I started with antirez' DwarfStar[1] on one spark and that (~11-14tok/s generation, ~300-400 tok/s prompt processing) was enough of a taste for me to jump into 2 sparks, running the native quant of DSv4 Flash. Now at 40-50tok/s generation and ~2000 tok/s prefill with a model that I've seen reason through race conditions and be able to trivially pull off any straight-forward coding task, and remain coherent at 500k co", "context": "", "use_case": "local LLM inference on DGX Spark", "switched_from": "", "story": "Two Qwen3 models on one DGX Spark: the residency math"}]}, {"text": " and several say it is solid enough for daily coding and agent work.", "cites": [{"id": 799264304, "item_id": 49954019, "url": "https://news.ycombinator.com/item?id=49954019", "by": "kamranjon", "date": "2026-10-04", "stance": "recommend", "quote": "I use the q4 quant daily and it works really well.", "excerpt": "Dwarfstar already supports this, curious how it compares, but I use the q4 quant daily and it works really well. https://github.com/antirez/ds4/blob/main/docs/MODELS.md#qwen38-flash-next", "context": "", "use_case": "running Qwen locally", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 794038304, "item_id": 49627394, "url": "https://news.ycombinator.com/item?id=49627394", "by": "kamranjon", "date": "2026-09-09", "stance": "recommend", "quote": "I've been using ds4 for months now and it's already got support for the new vision model", "excerpt": "Hey there! I do the same but I use dwarfstar at a 2-bit quant: https://github.com/antirez/ds4 I'm curious if you've tried dwarfstar and decided to move to llama.cpp and 3 bit quants or what made you go that route instead? I've been using ds4 for months now and it's already got support for the new vision model, haven't tried it yet, still on 0731 but it's been very solid for me.", "context": "", "use_case": "running DeepSeek V4 locally at 2-bit quant", "switched_from": "", "story": "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro"}, {"id": 776708800, "item_id": 48544300, "url": "https://news.ycombinator.com/item?id=48544300", "by": "Lwerewolf", "date": "2026-06-15", "stance": "recommend", "quote": "it absolutely is usable", "excerpt": "mbp16 m5 max 128gb, antirez/ds4, deepseekv4-flash. Works well for relatively dense (let's say <20k LoC per project) C codebases that are essentially a bunch of custom specialized stores, http servers, network infra, media transformers, etc. Runs through Pi with a custom prompt (basically \"don't speculate blindly, isolate things, make them traceable and measurable, then verify\") and behind a pretty restrictive bwrap s", "context": "works on dense C codebases", "use_case": "local coding inference", "switched_from": "", "story": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?"}]}, {"text": " The most common caveats are slow generation on some setups", "cites": [{"id": 776173857, "item_id": 48510866, "url": "https://news.ycombinator.com/item?id=48510866", "by": "jumploops", "date": "2026-06-12", "stance": "recommend", "quote": "running via antirez's ds4", "excerpt": "I've been quite impressed with DeepSeek v4 Flash running via antirez's ds4[0]. It feels like a GPT-4 class model in terms of \"stored knowledge\" but is better at long-horizon tool calling than any of the GPT-4 class models. Running on a 128GB MBP M4 Max, I'm getting ~24 t/s on generation and ~200 t/s on prefill. I was expecting it to feel slow, and it certainly does when e.g. generating code, but it's surprisingly use", "context": "128GB MBP M4 Max", "use_case": "running DeepSeek V4 Flash locally", "switched_from": "", "story": "How to Setup a Local Coding Agent on macOS"}, {"id": 779260672, "item_id": 48703792, "url": "https://news.ycombinator.com/item?id=48703792", "by": "pixelpoet", "date": "2026-06-27", "stance": "recommend", "quote": "have been extremely excited about Antirez's (Redis author) work on DS4", "excerpt": "I have two 128gb Strix Halos and have been extremely excited about Antirez's (Redis author) work on DS4, especially with 4bit quant using two machines: https://github.com/antirez/ds4 Right now the speed isn't good for GLM 5.2, Deepseek V4 Flash speed is okay for me (actually reading the output) and quite usable. See kyuz0's great recent video here: https://www.youtube.com/watch?v=PkKXm_mKCCM With a bit more speed and", "context": "owns two Strix Halo machines", "use_case": "local LLM inference on Strix Halo", "switched_from": "", "story": "AMD Strix Halo RDMA Cluster Setup Guide"}, {"id": 793898081, "item_id": 49618630, "url": "https://news.ycombinator.com/item?id=49618630", "by": "D13Fd", "date": "2026-09-08", "stance": "compare", "quote": "I\u2019ve been using 27B more and more over DS4.", "excerpt": "I\u2019m running 27B on a 5090 as well, and the results have been really strong. It does almost as well as, and sometimes better than, a 121gb DS4 model running on an M5 Max 128gb. 27B also flies on the 5090, and at medium think it returns results many times faster than my DS4 setup (the default xhigh is basically broken, though). For the kinds of things I use a local model for (legal document review), it\u2019s just spectacul", "context": "", "use_case": "local legal document review", "switched_from": "", "story": "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses"}]}, {"text": " and fewer features and supported models than llama.cpp.", "cites": [{"id": 794057473, "item_id": 49628592, "url": "https://news.ycombinator.com/item?id=49628592", "by": "tarruda", "date": "2026-09-09", "stance": "compare", "quote": "I didn't see anything that would make me want to stick with it vs llama.cpp.", "excerpt": "I tried dwarfstar when llama.cpp DSV4 support was still very weak, and while it worked, I didn't see anything that would make me want to stick with it vs llama.cpp. llama.cpp is simply better with its awesome built-in webui, router and server APIs and certainly support much more models and quantizations than dwarfstar. Since then, I started maintaining my own vibe coded dsv4 branch with metal optimizations, so I actu", "context": "", "use_case": "running DeepSeek V4 locally", "switched_from": "", "story": "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro"}]}], "praise": [{"label": "Large local models", "summary": "Lets 128GB machines run big sparse models like DeepSeek V4 Flash locally.", "quotes": [{"id": 779590305, "item_id": 48724394, "url": "https://news.ycombinator.com/item?id=48724394", "by": "ljosifov", "date": "2026-06-29", "stance": "recommend", "quote": "DeepSeek-V4-Flash will fit, served by DwarfStar", "excerpt": "Running 27B dense model on M5 128GB is ok, but one can do better. On M5 128GB one can make use of the ram and use sparse MoE. For example, DeepSeek-V4-Flash will fit, served by DwarfStar (https://github.com/antirez/ds4). One will probably improve 2x the token/sec speed, given DS4F 13B activated params in the MoE are ~1/2 of the ~27B of the dense Qwen. 27B Of the Qwen fit even on a cheaper 24GB card, e.g. amd 7900xtx ", "context": "", "use_case": "serving DeepSeek V4 Flash locally", "switched_from": "", "story": "Qwen 3.6 27B is the sweet spot for local development"}, {"id": 776173857, "item_id": 48510866, "url": "https://news.ycombinator.com/item?id=48510866", "by": "jumploops", "date": "2026-06-12", "stance": "recommend", "quote": "running via antirez's ds4", "excerpt": "I've been quite impressed with DeepSeek v4 Flash running via antirez's ds4[0]. It feels like a GPT-4 class model in terms of \"stored knowledge\" but is better at long-horizon tool calling than any of the GPT-4 class models. Running on a 128GB MBP M4 Max, I'm getting ~24 t/s on generation and ~200 t/s on prefill. I was expecting it to feel slow, and it certainly does when e.g. generating code, but it's surprisingly use", "context": "128GB MBP M4 Max", "use_case": "running DeepSeek V4 Flash locally", "switched_from": "", "story": "How to Setup a Local Coding Agent on macOS"}, {"id": 777514996, "item_id": 48594687, "url": "https://news.ycombinator.com/item?id=48594687", "by": "shsh1312", "date": "2026-06-18", "stance": "recommend", "quote": "allow optimized local inference through open-source tools such as: https://github.com/antirez/ds4", "excerpt": "Claude-Code and Codex in combination, combined with an IDE such as Google Antigravity or VisualStudio-Code are very powerful tools, if your company can invest in hardware the new Mac Studio and MacBook Pro allow optimized local inference through open-source tools such as: https://github.com/antirez/ds4", "context": "", "use_case": "local inference", "switched_from": "", "story": "Ask HN: What tools are you using for AI-assisted code review?"}]}, {"label": "Daily reliability", "summary": "Users report it working well and staying solid in daily use.", "quotes": [{"id": 799264304, "item_id": 49954019, "url": "https://news.ycombinator.com/item?id=49954019", "by": "kamranjon", "date": "2026-10-04", "stance": "recommend", "quote": "I use the q4 quant daily and it works really well.", "excerpt": "Dwarfstar already supports this, curious how it compares, but I use the q4 quant daily and it works really well. https://github.com/antirez/ds4/blob/main/docs/MODELS.md#qwen38-flash-next", "context": "", "use_case": "running Qwen locally", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 794038304, "item_id": 49627394, "url": "https://news.ycombinator.com/item?id=49627394", "by": "kamranjon", "date": "2026-09-09", "stance": "recommend", "quote": "I've been using ds4 for months now and it's already got support for the new vision model", "excerpt": "Hey there! I do the same but I use dwarfstar at a 2-bit quant: https://github.com/antirez/ds4 I'm curious if you've tried dwarfstar and decided to move to llama.cpp and 3 bit quants or what made you go that route instead? I've been using ds4 for months now and it's already got support for the new vision model, haven't tried it yet, still on 0731 but it's been very solid for me.", "context": "", "use_case": "running DeepSeek V4 locally at 2-bit quant", "switched_from": "", "story": "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro"}, {"id": 776708800, "item_id": 48544300, "url": "https://news.ycombinator.com/item?id=48544300", "by": "Lwerewolf", "date": "2026-06-15", "stance": "recommend", "quote": "it absolutely is usable", "excerpt": "mbp16 m5 max 128gb, antirez/ds4, deepseekv4-flash. Works well for relatively dense (let's say <20k LoC per project) C codebases that are essentially a bunch of custom specialized stores, http servers, network infra, media transformers, etc. Runs through Pi with a custom prompt (basically \"don't speculate blindly, isolate things, make them traceable and measurable, then verify\") and behind a pretty restrictive bwrap s", "context": "works on dense C codebases", "use_case": "local coding inference", "switched_from": "", "story": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?"}]}, {"label": "Model and features", "summary": "Supports DeepSeek 4, Qwen3.8 Flash Next, a vision model and runtime steering vectors.", "quotes": [{"id": 801086898, "item_id": 50067931, "url": "https://news.ycombinator.com/item?id=50067931", "by": "MadameMinty", "date": "2026-10-13", "stance": "recommend", "quote": "DeepSeek 4 family + more.", "excerpt": "Then check out those: https://github.com/prabanta-dev/janas <- this, focuses on Qwen3-Next-80B-A3B. https://github.com/Niko1221/Strata <- Qwen3.8-Flash-Next. I use this (tuned) and I cannot understate how satisfied I am. https://github.com/antirez/ds4 <- DeepSeek 4 family + more.", "context": "", "use_case": "running DeepSeek 4 locally", "switched_from": "", "story": "Janas: Run LLMs larger than RAM by streaming experts from SSD"}, {"id": 794038304, "item_id": 49627394, "url": "https://news.ycombinator.com/item?id=49627394", "by": "kamranjon", "date": "2026-09-09", "stance": "recommend", "quote": "I've been using ds4 for months now and it's already got support for the new vision model", "excerpt": "Hey there! I do the same but I use dwarfstar at a 2-bit quant: https://github.com/antirez/ds4 I'm curious if you've tried dwarfstar and decided to move to llama.cpp and 3 bit quants or what made you go that route instead? I've been using ds4 for months now and it's already got support for the new vision model, haven't tried it yet, still on 0731 but it's been very solid for me.", "context": "", "use_case": "running DeepSeek V4 locally at 2-bit quant", "switched_from": "", "story": "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro"}, {"id": 796441504, "item_id": 49777594, "url": "https://news.ycombinator.com/item?id=49777594", "by": "wren6991", "date": "2026-09-20", "stance": "recommend", "quote": "Antirez's DS4 already supports this", "excerpt": "Speaking to the \"uncensored model\" angle: there's little reason to distribute abliterated weights anyway. Instead of orthogonalising the weights that write back to the residual stream, you can just orthogonalise the activations themselves. It's equivalent. Orthogonalising activations at runtime is computationally cheap. Just distribute the refusal vectors (few thousand floats per layer), then run against the stock we", "context": "", "use_case": "refusal steering at runtime", "switched_from": "", "story": "Pirate Face Rescues LLM Models from Deletion"}]}, {"label": "Usable speed", "summary": "Commenters report usable token rates on Spark, Strix Halo and M4 Max hardware.", "quotes": [{"id": 777910720, "item_id": 48619420, "url": "https://news.ycombinator.com/item?id=48619420", "by": "wolttam", "date": "2026-06-21", "stance": "recommend", "quote": "I started with antirez' DwarfStar[1] on one spark and that (~11-14tok/s generation, ~300-400 tok/s prompt processing) was enough of a taste for me", "excerpt": "I started with antirez' DwarfStar[1] on one spark and that (~11-14tok/s generation, ~300-400 tok/s prompt processing) was enough of a taste for me to jump into 2 sparks, running the native quant of DSv4 Flash. Now at 40-50tok/s generation and ~2000 tok/s prefill with a model that I've seen reason through race conditions and be able to trivially pull off any straight-forward coding task, and remain coherent at 500k co", "context": "", "use_case": "local LLM inference on DGX Spark", "switched_from": "", "story": "Two Qwen3 models on one DGX Spark: the residency math"}, {"id": 779260672, "item_id": 48703792, "url": "https://news.ycombinator.com/item?id=48703792", "by": "pixelpoet", "date": "2026-06-27", "stance": "recommend", "quote": "have been extremely excited about Antirez's (Redis author) work on DS4", "excerpt": "I have two 128gb Strix Halos and have been extremely excited about Antirez's (Redis author) work on DS4, especially with 4bit quant using two machines: https://github.com/antirez/ds4 Right now the speed isn't good for GLM 5.2, Deepseek V4 Flash speed is okay for me (actually reading the output) and quite usable. See kyuz0's great recent video here: https://www.youtube.com/watch?v=PkKXm_mKCCM With a bit more speed and", "context": "owns two Strix Halo machines", "use_case": "local LLM inference on Strix Halo", "switched_from": "", "story": "AMD Strix Halo RDMA Cluster Setup Guide"}, {"id": 776173857, "item_id": 48510866, "url": "https://news.ycombinator.com/item?id=48510866", "by": "jumploops", "date": "2026-06-12", "stance": "recommend", "quote": "running via antirez's ds4", "excerpt": "I've been quite impressed with DeepSeek v4 Flash running via antirez's ds4[0]. It feels like a GPT-4 class model in terms of \"stored knowledge\" but is better at long-horizon tool calling than any of the GPT-4 class models. Running on a 128GB MBP M4 Max, I'm getting ~24 t/s on generation and ~200 t/s on prefill. I was expecting it to feel slow, and it certainly does when e.g. generating code, but it's surprisingly use", "context": "128GB MBP M4 Max", "use_case": "running DeepSeek V4 Flash locally", "switched_from": "", "story": "How to Setup a Local Coding Agent on macOS"}]}, {"label": "Focused design", "summary": "Praised for targeting the main use cases instead of every possible setup.", "quotes": [{"id": 799117616, "item_id": 49944851, "url": "https://news.ycombinator.com/item?id=49944851", "by": "westoque", "date": "2026-10-03", "stance": "recommend", "quote": "this is such a great quote about building software in the AI age", "excerpt": "from the project https://github.com/antirez/ds4 it was said that: > It must be more a working template for the biggest use cases, without trying to cover every possible setup. this is such a great quote about building software in the AI age. given that everyone has a different use case, the best way for open source software to be built is to build it for the general use case and specialization could be done afterward", "context": "", "use_case": "running LLMs locally", "switched_from": "", "story": "From the creator of Redis; run LLM locally with ds4"}]}, {"label": "Subscription alternative", "summary": "Cited as a way to get capable local models without expensive AI subscriptions.", "quotes": [{"id": 777515252, "item_id": 48594703, "url": "https://news.ycombinator.com/item?id=48594703", "by": "shsh1312", "date": "2026-06-18", "stance": "recommend", "quote": "Antirez/DS4 is an example", "excerpt": "Just a minimum, and this has allowed me to optimize my way of programming to the maximum, I have written an entire operating system (\"https://github.com/olmox001/NexsOS1\") and I am continuing to maintain it practically only with Claude's $20 plan and Codex and Antigravity's free plans, if I had more to invest I would certainly be able to do better, But I find super subscriptions excessive, there are local inference c", "context": "low-budget AI coding", "use_case": "local inference", "switched_from": "", "story": "How much $ you spend for AI to code?"}]}], "caveats": [{"label": "Generation speed", "summary": "Generation can feel slow, and some models such as GLM 5.2 run poorly.", "quotes": [{"id": 776173857, "item_id": 48510866, "url": "https://news.ycombinator.com/item?id=48510866", "by": "jumploops", "date": "2026-06-12", "stance": "recommend", "quote": "running via antirez's ds4", "excerpt": "I've been quite impressed with DeepSeek v4 Flash running via antirez's ds4[0]. It feels like a GPT-4 class model in terms of \"stored knowledge\" but is better at long-horizon tool calling than any of the GPT-4 class models. Running on a 128GB MBP M4 Max, I'm getting ~24 t/s on generation and ~200 t/s on prefill. I was expecting it to feel slow, and it certainly does when e.g. generating code, but it's surprisingly use", "context": "128GB MBP M4 Max", "use_case": "running DeepSeek V4 Flash locally", "switched_from": "", "story": "How to Setup a Local Coding Agent on macOS"}, {"id": 779260672, "item_id": 48703792, "url": "https://news.ycombinator.com/item?id=48703792", "by": "pixelpoet", "date": "2026-06-27", "stance": "recommend", "quote": "have been extremely excited about Antirez's (Redis author) work on DS4", "excerpt": "I have two 128gb Strix Halos and have been extremely excited about Antirez's (Redis author) work on DS4, especially with 4bit quant using two machines: https://github.com/antirez/ds4 Right now the speed isn't good for GLM 5.2, Deepseek V4 Flash speed is okay for me (actually reading the output) and quite usable. See kyuz0's great recent video here: https://www.youtube.com/watch?v=PkKXm_mKCCM With a bit more speed and", "context": "owns two Strix Halo machines", "use_case": "local LLM inference on Strix Halo", "switched_from": "", "story": "AMD Strix Halo RDMA Cluster Setup Guide"}, {"id": 793898081, "item_id": 49618630, "url": "https://news.ycombinator.com/item?id=49618630", "by": "D13Fd", "date": "2026-09-08", "stance": "compare", "quote": "I\u2019ve been using 27B more and more over DS4.", "excerpt": "I\u2019m running 27B on a 5090 as well, and the results have been really strong. It does almost as well as, and sometimes better than, a 121gb DS4 model running on an M5 Max 128gb. 27B also flies on the 5090, and at medium think it returns results many times faster than my DS4 setup (the default xhigh is basically broken, though). For the kinds of things I use a local model for (legal document review), it\u2019s just spectacul", "context": "", "use_case": "local legal document review", "switched_from": "", "story": "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses"}]}, {"label": "Versus llama.cpp", "summary": "One user found llama.cpp better for its web UI, router, APIs and broader model support.", "quotes": [{"id": 794057473, "item_id": 49628592, "url": "https://news.ycombinator.com/item?id=49628592", "by": "tarruda", "date": "2026-09-09", "stance": "compare", "quote": "I didn't see anything that would make me want to stick with it vs llama.cpp.", "excerpt": "I tried dwarfstar when llama.cpp DSV4 support was still very weak, and while it worked, I didn't see anything that would make me want to stick with it vs llama.cpp. llama.cpp is simply better with its awesome built-in webui, router and server APIs and certainly support much more models and quantizations than dwarfstar. Since then, I started maintaining my own vibe coded dsv4 branch with metal optimizations, so I actu", "context": "", "use_case": "running DeepSeek V4 locally", "switched_from": "", "story": "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro"}]}, {"label": "Strix Halo stability", "summary": "One commenter is waiting for it to stabilize on Strix Halo.", "quotes": [{"id": 776701744, "item_id": 48543859, "url": "https://news.ycombinator.com/item?id=48543859", "by": "christkv", "date": "2026-06-15", "stance": "neutral", "quote": "Waiting for this https://github.com/antirez/ds4 to stabilize for strix halo.", "excerpt": "Waiting for this https://github.com/antirez/ds4 to stabilize for strix halo.", "context": "", "use_case": "local inference on Strix Halo", "switched_from": "", "story": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?"}]}, {"label": "Hardware cost", "summary": "Needs costly high-memory hardware, and one user says running locally makes no economic sense.", "quotes": [{"id": 788661696, "item_id": 49291356, "url": "https://news.ycombinator.com/item?id=49291356", "by": "Muromec", "date": "2026-08-13", "stance": "neutral", "quote": "also use DS4", "excerpt": "I agree with your overall and also use DS4. Economically it makes no sense to run locally (electricity alone would be more expensive than what I pay our brothers in communist faith). But as a matter of fact I do have a tinkering urge and a 3d printer, so I better stay away from this stuff.", "context": "has a 3D printer, likes tinkering", "use_case": "local LLM inference", "switched_from": "", "story": "We Have AI at Home Chapter 1: A Box of Scraps"}, {"id": 779260672, "item_id": 48703792, "url": "https://news.ycombinator.com/item?id=48703792", "by": "pixelpoet", "date": "2026-06-27", "stance": "recommend", "quote": "have been extremely excited about Antirez's (Redis author) work on DS4", "excerpt": "I have two 128gb Strix Halos and have been extremely excited about Antirez's (Redis author) work on DS4, especially with 4bit quant using two machines: https://github.com/antirez/ds4 Right now the speed isn't good for GLM 5.2, Deepseek V4 Flash speed is okay for me (actually reading the output) and quite usable. See kyuz0's great recent video here: https://www.youtube.com/watch?v=PkKXm_mKCCM With a bit more speed and", "context": "owns two Strix Halo machines", "use_case": "local LLM inference on Strix Halo", "switched_from": "", "story": "AMD Strix Halo RDMA Cluster Setup Guide"}]}], "who": [{"by": "GTP", "context": "has not tried it personally", "url": "https://news.ycombinator.com/item?id=49656255", "date": "2026-09-11"}, {"by": "Muromec", "context": "has a 3D printer, likes tinkering", "url": "https://news.ycombinator.com/item?id=49291356", "date": "2026-08-13"}, {"by": "pixelpoet", "context": "owns two Strix Halo machines", "url": "https://news.ycombinator.com/item?id=48703792", "date": "2026-06-27"}, {"by": "kamranjon", "context": "codes Go and Typescript on a 128GB MacBook Pro", "url": "https://news.ycombinator.com/item?id=48680104", "date": "2026-06-25"}, {"by": "shsh1312", "context": "low-budget AI coding", "url": "https://news.ycombinator.com/item?id=48594703", "date": "2026-06-18"}, {"by": "Lwerewolf", "context": "works on dense C codebases", "url": "https://news.ycombinator.com/item?id=48544300", "date": "2026-06-15"}, {"by": "jumploops", "context": "128GB MBP M4 Max", "url": "https://news.ycombinator.com/item?id=48510866", "date": "2026-06-12"}], "tags": ["for local inference", "for Apple Silicon Macs", "for Strix Halo users", "for coding agents", "for private offline use"], "use_cases": ["local llm inference", "local inference", "running deepseek v4 locally", "running deepseek v4 flash locally", "running local models", "running deepseek 4 locally"], "compared": [{"slug": "deepseek-v4-flash", "name": "DeepSeek V4 Flash", "co": 7, "switched_from": 0, "switched_to": 0}, {"slug": "qwen-3-8", "name": "Qwen 3.8", "co": 3, "switched_from": 0, "switched_to": 1}, {"slug": "llama-cpp", "name": "llama.cpp", "co": 3, "switched_from": 0, "switched_to": 1}, {"slug": "deepseek", "name": "DeepSeek", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "glm", "name": "GLM", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "antigravity", "name": "Antigravity", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "codex", "name": "Codex", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "pi", "name": "Pi", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [{"id": 799318000, "item_id": 49957375, "url": "https://news.ycombinator.com/item?id=49957375", "by": "AntiRush", "date": "2026-10-04", "stance": "recommend", "quote": "The ds4 q4 quant performs a lot better than other similar sizes that I've seen.", "excerpt": "I've been working on support for this model in ds4 on the RTX 6000 pro - it's been really great for my use cases. The ds4 q4 quant performs a lot better than other similar sizes that I've seen. Using the Q4 quant on an RTX 6000 Pro Workstation Edition at 450 watts: Code: prefill 1,251 tok/s decode 255.26 tok/s Prose: prefill 1,251 tok/s decode 198.78 tok/s Most important for me, I can run 4 concurrent streams at 400+", "context": "works on model support in ds4 on an RTX 6000 Pro", "use_case": "local LLM inference", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 799024609, "item_id": 49939038, "url": "https://news.ycombinator.com/item?id=49939038", "by": "neomantra", "date": "2026-10-02", "stance": "recommend", "quote": "I maintain a fork of ds4 as shared libraries", "excerpt": "I maintain a fork of ds4 as shared libraries and thus can be used with other languages via FFI, along with public builds/binaries [1]. I made ds4go [2] against ds4 using techniques inspired by yzma. In addition to the library bindings, we have a small library of tools (workspace for view/edit, scratchpad for persistence) and making your own is registering a Go function. And in recent weeks, I added the Vision and Qwe", "context": "maintains a ds4 fork", "use_case": "local LLM inference via FFI", "switched_from": "", "story": "From the creator of Redis; run LLM locally with ds4"}, {"id": 778319440, "item_id": 48644965, "url": "https://news.ycombinator.com/item?id=48644965", "by": "antirez", "date": "2026-06-23", "stance": "neutral", "quote": "DwarfStar work in progress numbers: I see 14 tokens/sec generation", "excerpt": "DwarfStar work in progress numbers: I see 14 tokens/sec generation, that slopes to 10 t/s with longer 10k or more context size. Consider that the indexed attention requires evaluating 2048 selected rows, 2x DeepSeek and with less compression, so the performances with larger contexts here to south faster. Prefill can be 180 t/s on small contexts to 150 t/s and less with larger contexts. I used DeepSeek v4 PRO in this ", "context": "", "use_case": "local LLM inference", "switched_from": "", "story": "Unsloth GLM-5.2 \u2013 How to Run Locally"}], "threads": [{"url": "https://news.ycombinator.com/item?id=49936575", "title": "From the creator of Redis; run LLM locally with ds4", "mentions": 4}, {"url": "https://news.ycombinator.com/item?id=48542100", "title": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?", "mentions": 3}, {"url": "https://news.ycombinator.com/item?id=49953495", "title": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=49624603", "title": "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=50129045", "title": "DeepSeek-v4.1-Flash-1017", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=50063717", "title": "Janas: Run LLMs larger than RAM by streaming experts from SSD", "mentions": 1}], "sentiment_by_month": {"2025-11": {}, "2025-12": {}, "2026-01": {}, "2026-02": {}, "2026-03": {}, "2026-04": {}, "2026-05": {}, "2026-06": {"recommend": 8, "neutral": 1}, "2026-07": {"recommend": 2, "compare": 1}, "2026-08": {"neutral": 1}, "2026-09": {"compare": 2, "recommend": 3}, "2026-10": {"recommend": 4, "neutral": 1}}, "quotes": [{"id": 802092353, "item_id": 50130772, "url": "https://news.ycombinator.com/item?id=50130772", "by": "kamranjon", "date": "2026-10-17", "stance": "neutral", "quote": "Try the Qwen 3.8 flash next model on DwarfStar", "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": "running local models", "switched_from": "", "story": "DeepSeek-v4.1-Flash-1017"}, {"id": 801086898, "item_id": 50067931, "url": "https://news.ycombinator.com/item?id=50067931", "by": "MadameMinty", "date": "2026-10-13", "stance": "recommend", "quote": "DeepSeek 4 family + more.", "excerpt": "Then check out those: https://github.com/prabanta-dev/janas <- this, focuses on Qwen3-Next-80B-A3B. https://github.com/Niko1221/Strata <- Qwen3.8-Flash-Next. I use this (tuned) and I cannot understate how satisfied I am. https://github.com/antirez/ds4 <- DeepSeek 4 family + more.", "context": "", "use_case": "running DeepSeek 4 locally", "switched_from": "", "story": "Janas: Run LLMs larger than RAM by streaming experts from SSD"}, {"id": 799264304, "item_id": 49954019, "url": "https://news.ycombinator.com/item?id=49954019", "by": "kamranjon", "date": "2026-10-04", "stance": "recommend", "quote": "I use the q4 quant daily and it works really well.", "excerpt": "Dwarfstar already supports this, curious how it compares, but I use the q4 quant daily and it works really well. https://github.com/antirez/ds4/blob/main/docs/MODELS.md#qwen38-flash-next", "context": "", "use_case": "running Qwen locally", "switched_from": "", "story": "Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s"}, {"id": 799117616, "item_id": 49944851, "url": "https://news.ycombinator.com/item?id=49944851", "by": "westoque", "date": "2026-10-03", "stance": "recommend", "quote": "this is such a great quote about building software in the AI age", "excerpt": "from the project https://github.com/antirez/ds4 it was said that: > It must be more a working template for the biggest use cases, without trying to cover every possible setup. this is such a great quote about building software in the AI age. given that everyone has a different use case, the best way for open source software to be built is to build it for the general use case and specialization could be done afterward", "context": "", "use_case": "running LLMs locally", "switched_from": "", "story": "From the creator of Redis; run LLM locally with ds4"}, {"id": 799013728, "item_id": 49938358, "url": "https://news.ycombinator.com/item?id=49938358", "by": "twoodfin", "date": "2026-10-02", "stance": "neutral", "quote": "The project GitHub page is a much better introduction for the hn crowd.", "excerpt": "https://github.com/antirez/ds4 The project GitHub page is a much better introduction for the hn crowd.", "context": "", "use_case": "run LLM locally", "switched_from": "", "story": "From the creator of Redis; run LLM locally with ds4"}, {"id": 799012384, "item_id": 49938274, "url": "https://news.ycombinator.com/item?id=49938274", "by": "vlowther", "date": "2026-10-02", "stance": "recommend", "quote": "It is pretty nifty.", "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": "local LLM inference", "switched_from": "", "story": "From the creator of Redis; run LLM locally with ds4"}, {"id": 796441504, "item_id": 49777594, "url": "https://news.ycombinator.com/item?id=49777594", "by": "wren6991", "date": "2026-09-20", "stance": "recommend", "quote": "Antirez's DS4 already supports this", "excerpt": "Speaking to the \"uncensored model\" angle: there's little reason to distribute abliterated weights anyway. Instead of orthogonalising the weights that write back to the residual stream, you can just orthogonalise the activations themselves. It's equivalent. Orthogonalising activations at runtime is computationally cheap. Just distribute the refusal vectors (few thousand floats per layer), then run against the stock we", "context": "", "use_case": "refusal steering at runtime", "switched_from": "", "story": "Pirate Face Rescues LLM Models from Deletion"}, {"id": 794500080, "item_id": 49656255, "url": "https://news.ycombinator.com/item?id=49656255", "by": "GTP", "date": "2026-09-11", "stance": "recommend", "quote": "it seems that this project is much more than PoC and actually both usable and useful for people doing local inference", "excerpt": "I also struggle to reconcile these things, I guess the only way would be to try and see if it works for me, without caring too much if it works for other people on the internet. I would just like to add that, DwarfStar's author (Salvatore Sanfilippo) is a strong supporter of the idea that you shouldn't read the code, and he says he never read DwarfStar's code. And still, it seems that this project is much more than P", "context": "has not tried it personally", "use_case": "local model inference", "switched_from": "", "story": "Astra for Coding: Why Are We Doing This Again?"}, {"id": 794057473, "item_id": 49628592, "url": "https://news.ycombinator.com/item?id=49628592", "by": "tarruda", "date": "2026-09-09", "stance": "compare", "quote": "I didn't see anything that would make me want to stick with it vs llama.cpp.", "excerpt": "I tried dwarfstar when llama.cpp DSV4 support was still very weak, and while it worked, I didn't see anything that would make me want to stick with it vs llama.cpp. llama.cpp is simply better with its awesome built-in webui, router and server APIs and certainly support much more models and quantizations than dwarfstar. Since then, I started maintaining my own vibe coded dsv4 branch with metal optimizations, so I actu", "context": "", "use_case": "running DeepSeek V4 locally", "switched_from": "", "story": "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro"}, {"id": 794038304, "item_id": 49627394, "url": "https://news.ycombinator.com/item?id=49627394", "by": "kamranjon", "date": "2026-09-09", "stance": "recommend", "quote": "I've been using ds4 for months now and it's already got support for the new vision model", "excerpt": "Hey there! I do the same but I use dwarfstar at a 2-bit quant: https://github.com/antirez/ds4 I'm curious if you've tried dwarfstar and decided to move to llama.cpp and 3 bit quants or what made you go that route instead? I've been using ds4 for months now and it's already got support for the new vision model, haven't tried it yet, still on 0731 but it's been very solid for me.", "context": "", "use_case": "running DeepSeek V4 locally at 2-bit quant", "switched_from": "", "story": "DeepSeek launching v4.1 flash cheaper and more capable than v4 pro"}, {"id": 793898081, "item_id": 49618630, "url": "https://news.ycombinator.com/item?id=49618630", "by": "D13Fd", "date": "2026-09-08", "stance": "compare", "quote": "I\u2019ve been using 27B more and more over DS4.", "excerpt": "I\u2019m running 27B on a 5090 as well, and the results have been really strong. It does almost as well as, and sometimes better than, a 121gb DS4 model running on an M5 Max 128gb. 27B also flies on the 5090, and at medium think it returns results many times faster than my DS4 setup (the default xhigh is basically broken, though). For the kinds of things I use a local model for (legal document review), it\u2019s just spectacul", "context": "", "use_case": "local legal document review", "switched_from": "", "story": "Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses"}, {"id": 788661696, "item_id": 49291356, "url": "https://news.ycombinator.com/item?id=49291356", "by": "Muromec", "date": "2026-08-13", "stance": "neutral", "quote": "also use DS4", "excerpt": "I agree with your overall and also use DS4. Economically it makes no sense to run locally (electricity alone would be more expensive than what I pay our brothers in communist faith). But as a matter of fact I do have a tinkering urge and a 3d printer, so I better stay away from this stuff.", "context": "has a 3D printer, likes tinkering", "use_case": "local LLM inference", "switched_from": "", "story": "We Have AI at Home Chapter 1: A Box of Scraps"}, {"id": 783754528, "item_id": 48984658, "url": "https://news.ycombinator.com/item?id=48984658", "by": "c4pt0r", "date": "2026-07-20", "stance": "recommend", "quote": "why not running DeepSeek V4 on top of https://github.com/antirez/ds4. I think the results could be really good.", "excerpt": "I'm a bit curious why not running DeepSeek V4 on top of https://github.com/antirez/ds4. I think the results could be really good.", "context": "", "use_case": "running DeepSeek V4 locally", "switched_from": "", "story": "Nativ: Run frontier open models locally on your Mac"}, {"id": 783586528, "item_id": 48974158, "url": "https://news.ycombinator.com/item?id=48974158", "by": "para_parolu", "date": "2026-07-19", "stance": "compare", "quote": "I wonder how it\u2019s compared to https://github.com/antirez/ds4", "excerpt": "I wonder how it\u2019s compared to https://github.com/antirez/ds4", "context": "", "use_case": "local LLM inference", "switched_from": "", "story": "6.4x faster than llama.cpp, 3.9x faster than MLX"}, {"id": 782281552, "item_id": 48892597, "url": "https://news.ycombinator.com/item?id=48892597", "by": "kamranjon", "date": "2026-07-13", "stance": "recommend", "quote": "The DS4 repo is a really interesting place to watch folks who heavily code with agents collaborate together in a way that seems pretty effective.", "excerpt": "Some of the most exciting engineering work is happening in the DS4 repo - and I'm watching it almost like a sports game. When the DSpark paper came out[1] the next day we had folks attempting to implement, working together, validating their failures. Eventually their work being synthesized into a PR[1] that admits performance is not ideal. Something antirez alluded to in one of his videos (speculative decoding is a g", "context": "", "use_case": "local inference", "switched_from": "", "story": "Control the Ideas, Not the Code"}, {"id": 779590305, "item_id": 48724394, "url": "https://news.ycombinator.com/item?id=48724394", "by": "ljosifov", "date": "2026-06-29", "stance": "recommend", "quote": "DeepSeek-V4-Flash will fit, served by DwarfStar", "excerpt": "Running 27B dense model on M5 128GB is ok, but one can do better. On M5 128GB one can make use of the ram and use sparse MoE. For example, DeepSeek-V4-Flash will fit, served by DwarfStar (https://github.com/antirez/ds4). One will probably improve 2x the token/sec speed, given DS4F 13B activated params in the MoE are ~1/2 of the ~27B of the dense Qwen. 27B Of the Qwen fit even on a cheaper 24GB card, e.g. amd 7900xtx ", "context": "", "use_case": "serving DeepSeek V4 Flash locally", "switched_from": "", "story": "Qwen 3.6 27B is the sweet spot for local development"}, {"id": 779260672, "item_id": 48703792, "url": "https://news.ycombinator.com/item?id=48703792", "by": "pixelpoet", "date": "2026-06-27", "stance": "recommend", "quote": "have been extremely excited about Antirez's (Redis author) work on DS4", "excerpt": "I have two 128gb Strix Halos and have been extremely excited about Antirez's (Redis author) work on DS4, especially with 4bit quant using two machines: https://github.com/antirez/ds4 Right now the speed isn't good for GLM 5.2, Deepseek V4 Flash speed is okay for me (actually reading the output) and quite usable. See kyuz0's great recent video here: https://www.youtube.com/watch?v=PkKXm_mKCCM With a bit more speed and", "context": "owns two Strix Halo machines", "use_case": "local LLM inference on Strix Halo", "switched_from": "", "story": "AMD Strix Halo RDMA Cluster Setup Guide"}, {"id": 778881664, "item_id": 48680104, "url": "https://news.ycombinator.com/item?id=48680104", "by": "kamranjon", "date": "2026-06-25", "stance": "recommend", "quote": "I use it as my main coding agent - so its running DS4 server on my 128gb mbp", "excerpt": "I use it as my main coding agent - so its running DS4 server on my 128gb mbp and I run the pi coding agent on my other machine which calls out to it. Mostly Go and Typescript work. I also use it in local agent mode if im coding directly on the machine which is nice cause you can save sessions and resume them, and so for personal projects and training related stuff it's been great. Even got an autoresearch loop going ", "context": "codes Go and Typescript on a 128GB MacBook Pro", "use_case": "local coding agent model server", "switched_from": "", "story": "Apple announces significant price increases for MacBooks, iPads, more"}, {"id": 778165281, "item_id": 48635330, "url": "https://news.ycombinator.com/item?id=48635330", "by": "sleepyeldrazi", "date": "2026-06-22", "stance": "recommend", "quote": "Inspired by [https://github.com/antirez/ds4](ds4)", "excerpt": "Inspired by [https://github.com/antirez/ds4](ds4), [https://github.com/CerebrasResearch/reap](REAP) and [https://huggingface.co/0xSero/DeepSeek-V4-Flash-162B](OxSero's Deepseek v4 reap) I wanted to push and see how much performance can be extracted from a single DGX Spark. It should also work day one (hopefully) on the upcoming Spark devices. I made three versions, one with 128 experts kept, one with 150 and the bigg", "context": "", "use_case": "running DeepSeek V4 Flash locally", "switched_from": "", "story": "DeepSeek V4 Flash optimized framework and model variants for DGX Spark"}, {"id": 777910720, "item_id": 48619420, "url": "https://news.ycombinator.com/item?id=48619420", "by": "wolttam", "date": "2026-06-21", "stance": "recommend", "quote": "I started with antirez' DwarfStar[1] on one spark and that (~11-14tok/s generation, ~300-400 tok/s prompt processing) was enough of a taste for me", "excerpt": "I started with antirez' DwarfStar[1] on one spark and that (~11-14tok/s generation, ~300-400 tok/s prompt processing) was enough of a taste for me to jump into 2 sparks, running the native quant of DSv4 Flash. Now at 40-50tok/s generation and ~2000 tok/s prefill with a model that I've seen reason through race conditions and be able to trivially pull off any straight-forward coding task, and remain coherent at 500k co", "context": "", "use_case": "local LLM inference on DGX Spark", "switched_from": "", "story": "Two Qwen3 models on one DGX Spark: the residency math"}, {"id": 777516544, "item_id": 48594784, "url": "https://news.ycombinator.com/item?id=48594784", "by": "shsh1312", "date": "2026-06-18", "stance": "recommend", "quote": "", "excerpt": "https://github.com/antirez/ds4", "context": "", "use_case": "local model for daily coding", "switched_from": "", "story": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?"}, {"id": 777515252, "item_id": 48594703, "url": "https://news.ycombinator.com/item?id=48594703", "by": "shsh1312", "date": "2026-06-18", "stance": "recommend", "quote": "Antirez/DS4 is an example", "excerpt": "Just a minimum, and this has allowed me to optimize my way of programming to the maximum, I have written an entire operating system (\"https://github.com/olmox001/NexsOS1\") and I am continuing to maintain it practically only with Claude's $20 plan and Codex and Antigravity's free plans, if I had more to invest I would certainly be able to do better, But I find super subscriptions excessive, there are local inference c", "context": "low-budget AI coding", "use_case": "local inference", "switched_from": "", "story": "How much $ you spend for AI to code?"}, {"id": 777514996, "item_id": 48594687, "url": "https://news.ycombinator.com/item?id=48594687", "by": "shsh1312", "date": "2026-06-18", "stance": "recommend", "quote": "allow optimized local inference through open-source tools such as: https://github.com/antirez/ds4", "excerpt": "Claude-Code and Codex in combination, combined with an IDE such as Google Antigravity or VisualStudio-Code are very powerful tools, if your company can invest in hardware the new Mac Studio and MacBook Pro allow optimized local inference through open-source tools such as: https://github.com/antirez/ds4", "context": "", "use_case": "local inference", "switched_from": "", "story": "Ask HN: What tools are you using for AI-assisted code review?"}, {"id": 776708800, "item_id": 48544300, "url": "https://news.ycombinator.com/item?id=48544300", "by": "Lwerewolf", "date": "2026-06-15", "stance": "recommend", "quote": "it absolutely is usable", "excerpt": "mbp16 m5 max 128gb, antirez/ds4, deepseekv4-flash. Works well for relatively dense (let's say <20k LoC per project) C codebases that are essentially a bunch of custom specialized stores, http servers, network infra, media transformers, etc. Runs through Pi with a custom prompt (basically \"don't speculate blindly, isolate things, make them traceable and measurable, then verify\") and behind a pretty restrictive bwrap s", "context": "works on dense C codebases", "use_case": "local coding inference", "switched_from": "", "story": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?"}, {"id": 776701744, "item_id": 48543859, "url": "https://news.ycombinator.com/item?id=48543859", "by": "christkv", "date": "2026-06-15", "stance": "neutral", "quote": "Waiting for this https://github.com/antirez/ds4 to stabilize for strix halo.", "excerpt": "Waiting for this https://github.com/antirez/ds4 to stabilize for strix halo.", "context": "", "use_case": "local inference on Strix Halo", "switched_from": "", "story": "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?"}, {"id": 776173857, "item_id": 48510866, "url": "https://news.ycombinator.com/item?id=48510866", "by": "jumploops", "date": "2026-06-12", "stance": "recommend", "quote": "running via antirez's ds4", "excerpt": "I've been quite impressed with DeepSeek v4 Flash running via antirez's ds4[0]. It feels like a GPT-4 class model in terms of \"stored knowledge\" but is better at long-horizon tool calling than any of the GPT-4 class models. Running on a 128GB MBP M4 Max, I'm getting ~24 t/s on generation and ~200 t/s on prefill. I was expecting it to feel slow, and it certainly does when e.g. generating code, but it's surprisingly use", "context": "128GB MBP M4 Max", "use_case": "running DeepSeek V4 Flash locally", "switched_from": "", "story": "How to Setup a Local Coding Agent on macOS"}]}