{"slug": "qwen-27b", "name": "Qwen 27B", "category": "ai", "category_name": "AI models and assistants", "description": "A mid-sized open-weight large language model from Alibaba's Qwen family, often run locally.", "recommenders": 8, "commenters": 11, "warners": 0, "score": 7.687, "rising": -1.29, "first": "2026-06-16", "latest": "2026-10-15", "why": "HN commenters like it for agentic work with consistent tool calls, everyday non-coding tasks, and private local use.", "why_is_quote": false, "has_verdict": true, "sentiment": {"recommend": 8, "neutral": 4, "compare": 0, "warn": 0}, "spark": [0, 0, 0, 0, 0, 0, 0, 2, 5, 1, 0, 0], "listed": true, "loved": ["Agentic tool use", "Model quality"], "caveat": "Speed and hardware", "short_quote": "I use this 27B for everyday tasks that aren\u2019t actually coding.", "icon": true, "domain": null, "github_repo": null, "official_url": "https://qwen.ai", "pricing": "free", "aliases": ["Qwen 27B", "Qwen-27B"], "verdict": [{"text": "HN commenters like it for agentic work with consistent tool calls,", "cites": [{"id": 786437664, "item_id": 49152354, "url": "https://news.ycombinator.com/item?id=49152354", "by": "wickedsight", "date": "2026-08-03", "stance": "recommend", "quote": "I use 27B in plan mode and 35B MoE in act mode.", "excerpt": "I use 27B in plan mode and 35B MoE in act mode. I noticed that is the best balance for me for consistent tool calls and intelligent planning. Takes some time to switch, but it's worth it for me.", "context": "", "use_case": "local planning", "switched_from": "", "story": "Qwen3.8-Max: A New Bar for Coding and Cowork"}, {"id": 781782097, "item_id": 48861381, "url": "https://news.ycombinator.com/item?id=48861381", "by": "verdverm", "date": "2026-07-10", "stance": "recommend", "quote": "I use this model for agentic work", "excerpt": "Unconvincing, I use this model for agentic work, and while it's obviously not as good, this person is using their personal setup and experience as a justification. 1. Mac is slow at prompt processing, the Spark is much better for coding. Most of the time in coding is spent in prompt processing. 2. DFlash addon improves tgen speed 2-3x. (https://huggingface.co/collections/z-lab/dflash) 3. Harness matters, they gave no", "context": "", "use_case": "agentic work", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}, {"id": 779633120, "item_id": 48727070, "url": "https://news.ycombinator.com/item?id=48727070", "by": "cpburns2009", "date": "2026-06-29", "stance": "recommend", "quote": "Qwen 27b works well enough in OpenCode though which is what I use.", "excerpt": "If Qwen is finetuned for a hardness, it'll be Qwen Code. Qwen 27b works well enough in OpenCode though which is what I use. My one complaint is it likes to get cute with bash commands instead of OpenCode's built-in tools. I use a skill to steer that.", "context": "", "use_case": "local coding agent", "switched_from": "", "story": "Qwen 3.6 27B is the sweet spot for local development"}]}, {"text": " everyday non-coding tasks,", "cites": [{"id": 781786080, "item_id": 48861630, "url": "https://news.ycombinator.com/item?id=48861630", "by": "edyos", "date": "2026-07-10", "stance": "recommend", "quote": "I use this 27B for everyday tasks that aren\u2019t actually coding.", "excerpt": "You\u2019re right about the coding, but not about the rest. I use this 27B for everyday tasks that aren\u2019t actually coding. When it comes to coding, GLM5.2 is, in my view, a marvel; it\u2019s inexpensive and works very well when used properly \u2013 but for everything else, from sorting emails to pre-generation and analysis of varying quality, it does the job that GPT and Claude can\u2019t do for free, and sometimes not quite as well. I ", "context": "", "use_case": "sorting emails, analysis", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}]}, {"text": " and private local use.", "cites": [{"id": 794167776, "item_id": 49635486, "url": "https://news.ycombinator.com/item?id=49635486", "by": "tompark", "date": "2026-09-09", "stance": "neutral", "quote": "except when using local models like Qwen-27B", "excerpt": "Yup. I'm more paranoid about AI training on my chats than most ppl are. While I use frontier models to do research and to build tools, I've tried to avoid discussing details about my projects (e.g. story worldbuilding) except when using local models like Qwen-27B. Of course this slows down progress and has a real opportunity cost, but it would kill all my motivation if I got \"scooped\" like that.", "context": "", "use_case": "private local chats", "switched_from": "", "story": "Culture Becomes a Dark Forest"}, {"id": 781786080, "item_id": 48861630, "url": "https://news.ycombinator.com/item?id=48861630", "by": "edyos", "date": "2026-07-10", "stance": "recommend", "quote": "I use this 27B for everyday tasks that aren\u2019t actually coding.", "excerpt": "You\u2019re right about the coding, but not about the rest. I use this 27B for everyday tasks that aren\u2019t actually coding. When it comes to coding, GLM5.2 is, in my view, a marvel; it\u2019s inexpensive and works very well when used properly \u2013 but for everything else, from sorting emails to pre-generation and analysis of varying quality, it does the job that GPT and Claude can\u2019t do for free, and sometimes not quite as well. I ", "context": "", "use_case": "sorting emails, analysis", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}]}, {"text": " The most common caveat is slow local inference or needing a better GPU,", "cites": [{"id": 777317776, "item_id": 48582361, "url": "https://news.ycombinator.com/item?id=48582361", "by": "asimovDev", "date": "2026-06-18", "stance": "neutral", "quote": "It succesfully identified how it worked even the obfuscated classes and methods.", "excerpt": "I used Qwen 27b 8 bit MLX version on a decompiled android APK recently. It succesfully identified how it worked even the obfuscated classes and methods. It wrote a 1000 line documentation with examples but the time was dreadful. At some point it slowed down to 5 t/s so the whole thing took over an hour , the writing of documentation alone was over 40 minutes, fans blasting the entire time.", "context": "ran 8-bit MLX version locally", "use_case": "analyzing decompiled Android APK", "switched_from": "", "story": "Local Qwen isn't a worse Opus, it's a different tool"}, {"id": 791703076, "item_id": 49481442, "url": "https://news.ycombinator.com/item?id=49481442", "by": "wolvoleo", "date": "2026-08-28", "stance": "neutral", "quote": "And I can't run qwen 27b reliably at the moment, I need a better gpu.", "excerpt": "Similar here, except I used Opus 4.8. I don't mind using cloud models to make code as long as I don't put actual private data into the model. And I can't run qwen 27b reliably at the moment, I need a better gpu. It was pretty easy mode like this tbh, though you do need to know what you're doing. I use it with llama-server and openwebui. I guess you can get it a lot more click to go with something like LM Studio thoug", "context": "", "use_case": "local model", "switched_from": "", "story": "Run Qwen3.8 27B locally: real numbers from my Mac Studio"}]}, {"text": " and it is not as strong as frontier models.", "cites": [{"id": 781782097, "item_id": 48861381, "url": "https://news.ycombinator.com/item?id=48861381", "by": "verdverm", "date": "2026-07-10", "stance": "recommend", "quote": "I use this model for agentic work", "excerpt": "Unconvincing, I use this model for agentic work, and while it's obviously not as good, this person is using their personal setup and experience as a justification. 1. Mac is slow at prompt processing, the Spark is much better for coding. Most of the time in coding is spent in prompt processing. 2. DFlash addon improves tgen speed 2-3x. (https://huggingface.co/collections/z-lab/dflash) 3. Harness matters, they gave no", "context": "", "use_case": "agentic work", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}, {"id": 781786080, "item_id": 48861630, "url": "https://news.ycombinator.com/item?id=48861630", "by": "edyos", "date": "2026-07-10", "stance": "recommend", "quote": "I use this 27B for everyday tasks that aren\u2019t actually coding.", "excerpt": "You\u2019re right about the coding, but not about the rest. I use this 27B for everyday tasks that aren\u2019t actually coding. When it comes to coding, GLM5.2 is, in my view, a marvel; it\u2019s inexpensive and works very well when used properly \u2013 but for everything else, from sorting emails to pre-generation and analysis of varying quality, it does the job that GPT and Claude can\u2019t do for free, and sometimes not quite as well. I ", "context": "", "use_case": "sorting emails, analysis", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}]}], "praise": [{"label": "Agentic tool use", "summary": "Handles agentic workflows and tool calls well in harnesses like OpenCode and MCP setups.", "quotes": [{"id": 786437664, "item_id": 49152354, "url": "https://news.ycombinator.com/item?id=49152354", "by": "wickedsight", "date": "2026-08-03", "stance": "recommend", "quote": "I use 27B in plan mode and 35B MoE in act mode.", "excerpt": "I use 27B in plan mode and 35B MoE in act mode. I noticed that is the best balance for me for consistent tool calls and intelligent planning. Takes some time to switch, but it's worth it for me.", "context": "", "use_case": "local planning", "switched_from": "", "story": "Qwen3.8-Max: A New Bar for Coding and Cowork"}, {"id": 781782097, "item_id": 48861381, "url": "https://news.ycombinator.com/item?id=48861381", "by": "verdverm", "date": "2026-07-10", "stance": "recommend", "quote": "I use this model for agentic work", "excerpt": "Unconvincing, I use this model for agentic work, and while it's obviously not as good, this person is using their personal setup and experience as a justification. 1. Mac is slow at prompt processing, the Spark is much better for coding. Most of the time in coding is spent in prompt processing. 2. DFlash addon improves tgen speed 2-3x. (https://huggingface.co/collections/z-lab/dflash) 3. Harness matters, they gave no", "context": "", "use_case": "agentic work", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}, {"id": 779633120, "item_id": 48727070, "url": "https://news.ycombinator.com/item?id=48727070", "by": "cpburns2009", "date": "2026-06-29", "stance": "recommend", "quote": "Qwen 27b works well enough in OpenCode though which is what I use.", "excerpt": "If Qwen is finetuned for a hardness, it'll be Qwen Code. Qwen 27b works well enough in OpenCode though which is what I use. My one complaint is it likes to get cute with bash commands instead of OpenCode's built-in tools. I use a skill to steer that.", "context": "", "use_case": "local coding agent", "switched_from": "", "story": "Qwen 3.6 27B is the sweet spot for local development"}]}, {"label": "Model quality", "summary": "Several commenters call it a great model, especially when grounded.", "quotes": [{"id": 784234753, "item_id": 49014672, "url": "https://news.ycombinator.com/item?id=49014672", "by": "InTheArena", "date": "2026-07-22", "stance": "recommend", "quote": "Qwen 27b is a great model", "excerpt": "I have a DGX and a Ryzen AI Max 395 - while I love both of them, there are a few critical things that leave the DGX in use, while the Ryzen \"just\" is my primary homelab server. The biggest thing is prefil numbers, and the performance impact of higher context sizes. Qwen 27b is a great model, nemotron is decent, gemma is workable. But all of them need reasonable context for reasonable outputs. Unfortunitly, as others ", "context": "owns a DGX and Ryzen AI Max 395", "use_case": "local LLM", "switched_from": "", "story": "Nvidia DGX Spark as a daily driver"}, {"id": 783488529, "item_id": 48968033, "url": "https://news.ycombinator.com/item?id=48968033", "by": "eurekin", "date": "2026-07-19", "stance": "recommend", "quote": "27b is a great model, if grounded.", "excerpt": "It's a mcp, so connects quite easily to agents. With mcpo, I also connected it to open-webui (which has better support for OpenAI style tools/functions). Used it in claude code with that mcp plugin set-up too. Only ever used it for managing homelab information, but it met initial expectations. 27b is a great model, if grounded. The query about physical hosts and routing... I haven't found a single hallucination (alto", "context": "", "use_case": "homelab information queries", "switched_from": "", "story": "Qwen3.8 is launching and going open-weight soon"}, {"id": 776968864, "item_id": 48560554, "url": "https://news.ycombinator.com/item?id=48560554", "by": "freehorse", "date": "2026-06-16", "stance": "recommend", "quote": "I have been using the qwen 27b and it is great", "excerpt": "> You have MoE models (gemma 26b, qwen 35b, north mini code 30b) who are pretty fast, but make a lot of mistakes This is sadly also my experience. I wish we had some MoE models with a higher ratio of active parameters per total. My experience is that the newer MoE models that can run in a 64b laptop have too few active parameters to be useful outside narrower, specific tasks. Mixtral 8x7b was a 14b active parameter (", "context": "", "use_case": "local inference on laptop", "switched_from": "", "story": "Running local models is good now"}]}, {"label": "Local and private", "summary": "Runs locally, so commenters use it for private projects and free everyday work.", "quotes": [{"id": 794167776, "item_id": 49635486, "url": "https://news.ycombinator.com/item?id=49635486", "by": "tompark", "date": "2026-09-09", "stance": "neutral", "quote": "except when using local models like Qwen-27B", "excerpt": "Yup. I'm more paranoid about AI training on my chats than most ppl are. While I use frontier models to do research and to build tools, I've tried to avoid discussing details about my projects (e.g. story worldbuilding) except when using local models like Qwen-27B. Of course this slows down progress and has a real opportunity cost, but it would kill all my motivation if I got \"scooped\" like that.", "context": "", "use_case": "private local chats", "switched_from": "", "story": "Culture Becomes a Dark Forest"}, {"id": 781786080, "item_id": 48861630, "url": "https://news.ycombinator.com/item?id=48861630", "by": "edyos", "date": "2026-07-10", "stance": "recommend", "quote": "I use this 27B for everyday tasks that aren\u2019t actually coding.", "excerpt": "You\u2019re right about the coding, but not about the rest. I use this 27B for everyday tasks that aren\u2019t actually coding. When it comes to coding, GLM5.2 is, in my view, a marvel; it\u2019s inexpensive and works very well when used properly \u2013 but for everything else, from sorting emails to pre-generation and analysis of varying quality, it does the job that GPT and Claude can\u2019t do for free, and sometimes not quite as well. I ", "context": "", "use_case": "sorting emails, analysis", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}, {"id": 777317776, "item_id": 48582361, "url": "https://news.ycombinator.com/item?id=48582361", "by": "asimovDev", "date": "2026-06-18", "stance": "neutral", "quote": "It succesfully identified how it worked even the obfuscated classes and methods.", "excerpt": "I used Qwen 27b 8 bit MLX version on a decompiled android APK recently. It succesfully identified how it worked even the obfuscated classes and methods. It wrote a 1000 line documentation with examples but the time was dreadful. At some point it slowed down to 5 t/s so the whole thing took over an hour , the writing of documentation alone was over 40 minutes, fans blasting the entire time.", "context": "ran 8-bit MLX version locally", "use_case": "analyzing decompiled Android APK", "switched_from": "", "story": "Local Qwen isn't a worse Opus, it's a different tool"}]}, {"label": "Everyday tasks", "summary": "Used for non-coding work such as sorting emails and analysis.", "quotes": [{"id": 781786080, "item_id": 48861630, "url": "https://news.ycombinator.com/item?id=48861630", "by": "edyos", "date": "2026-07-10", "stance": "recommend", "quote": "I use this 27B for everyday tasks that aren\u2019t actually coding.", "excerpt": "You\u2019re right about the coding, but not about the rest. I use this 27B for everyday tasks that aren\u2019t actually coding. When it comes to coding, GLM5.2 is, in my view, a marvel; it\u2019s inexpensive and works very well when used properly \u2013 but for everything else, from sorting emails to pre-generation and analysis of varying quality, it does the job that GPT and Claude can\u2019t do for free, and sometimes not quite as well. I ", "context": "", "use_case": "sorting emails, analysis", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}]}, {"label": "Code analysis", "summary": "Worked out how a decompiled, obfuscated Android APK functioned and documented it.", "quotes": [{"id": 777317776, "item_id": 48582361, "url": "https://news.ycombinator.com/item?id=48582361", "by": "asimovDev", "date": "2026-06-18", "stance": "neutral", "quote": "It succesfully identified how it worked even the obfuscated classes and methods.", "excerpt": "I used Qwen 27b 8 bit MLX version on a decompiled android APK recently. It succesfully identified how it worked even the obfuscated classes and methods. It wrote a 1000 line documentation with examples but the time was dreadful. At some point it slowed down to 5 t/s so the whole thing took over an hour , the writing of documentation alone was over 40 minutes, fans blasting the entire time.", "context": "ran 8-bit MLX version locally", "use_case": "analyzing decompiled Android APK", "switched_from": "", "story": "Local Qwen isn't a worse Opus, it's a different tool"}]}, {"label": "Distillation target", "summary": "A team distilled a larger model into Qwen 27B specialists with great success.", "quotes": [{"id": 786069953, "item_id": 49129372, "url": "https://news.ycombinator.com/item?id=49129372", "by": "teravor", "date": "2026-07-31", "stance": "recommend", "quote": "we used to distill GLM 5.2 into Qwen 27B specialists with great success.", "excerpt": "distilling K3 into DS4 flash will likely only be a good idea for specialists. the difference in model capacity is otherwise too large. we used to distill GLM 5.2 into Qwen 27B specialists with great success. developing the workflows is tricky though. we had the advantage of a straightforward mapping function in mind: English -> SQL which was relatively easy to pipeline training for.", "context": "team training specialist models", "use_case": "English to SQL specialist model", "switched_from": "", "story": "DeepSeek-V4-Flash Update"}]}], "caveats": [{"label": "Speed and hardware", "summary": "Local runs can be very slow, and some commenters need a better GPU to run it reliably.", "quotes": [{"id": 777317776, "item_id": 48582361, "url": "https://news.ycombinator.com/item?id=48582361", "by": "asimovDev", "date": "2026-06-18", "stance": "neutral", "quote": "It succesfully identified how it worked even the obfuscated classes and methods.", "excerpt": "I used Qwen 27b 8 bit MLX version on a decompiled android APK recently. It succesfully identified how it worked even the obfuscated classes and methods. It wrote a 1000 line documentation with examples but the time was dreadful. At some point it slowed down to 5 t/s so the whole thing took over an hour , the writing of documentation alone was over 40 minutes, fans blasting the entire time.", "context": "ran 8-bit MLX version locally", "use_case": "analyzing decompiled Android APK", "switched_from": "", "story": "Local Qwen isn't a worse Opus, it's a different tool"}, {"id": 791703076, "item_id": 49481442, "url": "https://news.ycombinator.com/item?id=49481442", "by": "wolvoleo", "date": "2026-08-28", "stance": "neutral", "quote": "And I can't run qwen 27b reliably at the moment, I need a better gpu.", "excerpt": "Similar here, except I used Opus 4.8. I don't mind using cloud models to make code as long as I don't put actual private data into the model. And I can't run qwen 27b reliably at the moment, I need a better gpu. It was pretty easy mode like this tbh, though you do need to know what you're doing. I use it with llama-server and openwebui. I guess you can get it a lot more click to go with something like LM Studio thoug", "context": "", "use_case": "local model", "switched_from": "", "story": "Run Qwen3.8 27B locally: real numbers from my Mac Studio"}, {"id": 784234753, "item_id": 49014672, "url": "https://news.ycombinator.com/item?id=49014672", "by": "InTheArena", "date": "2026-07-22", "stance": "recommend", "quote": "Qwen 27b is a great model", "excerpt": "I have a DGX and a Ryzen AI Max 395 - while I love both of them, there are a few critical things that leave the DGX in use, while the Ryzen \"just\" is my primary homelab server. The biggest thing is prefil numbers, and the performance impact of higher context sizes. Qwen 27b is a great model, nemotron is decent, gemma is workable. But all of them need reasonable context for reasonable outputs. Unfortunitly, as others ", "context": "owns a DGX and Ryzen AI Max 395", "use_case": "local LLM", "switched_from": "", "story": "Nvidia DGX Spark as a daily driver"}]}, {"label": "Behind frontier models", "summary": "Commenters say it is not as good as GPT, Claude, or GLM for coding.", "quotes": [{"id": 781782097, "item_id": 48861381, "url": "https://news.ycombinator.com/item?id=48861381", "by": "verdverm", "date": "2026-07-10", "stance": "recommend", "quote": "I use this model for agentic work", "excerpt": "Unconvincing, I use this model for agentic work, and while it's obviously not as good, this person is using their personal setup and experience as a justification. 1. Mac is slow at prompt processing, the Spark is much better for coding. Most of the time in coding is spent in prompt processing. 2. DFlash addon improves tgen speed 2-3x. (https://huggingface.co/collections/z-lab/dflash) 3. Harness matters, they gave no", "context": "", "use_case": "agentic work", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}, {"id": 781786080, "item_id": 48861630, "url": "https://news.ycombinator.com/item?id=48861630", "by": "edyos", "date": "2026-07-10", "stance": "recommend", "quote": "I use this 27B for everyday tasks that aren\u2019t actually coding.", "excerpt": "You\u2019re right about the coding, but not about the rest. I use this 27B for everyday tasks that aren\u2019t actually coding. When it comes to coding, GLM5.2 is, in my view, a marvel; it\u2019s inexpensive and works very well when used properly \u2013 but for everything else, from sorting emails to pre-generation and analysis of varying quality, it does the job that GPT and Claude can\u2019t do for free, and sometimes not quite as well. I ", "context": "", "use_case": "sorting emails, analysis", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}]}, {"label": "Needs grounding", "summary": "Commenters say it needs grounding and reasonable context for good outputs.", "quotes": [{"id": 783488529, "item_id": 48968033, "url": "https://news.ycombinator.com/item?id=48968033", "by": "eurekin", "date": "2026-07-19", "stance": "recommend", "quote": "27b is a great model, if grounded.", "excerpt": "It's a mcp, so connects quite easily to agents. With mcpo, I also connected it to open-webui (which has better support for OpenAI style tools/functions). Used it in claude code with that mcp plugin set-up too. Only ever used it for managing homelab information, but it met initial expectations. 27b is a great model, if grounded. The query about physical hosts and routing... I haven't found a single hallucination (alto", "context": "", "use_case": "homelab information queries", "switched_from": "", "story": "Qwen3.8 is launching and going open-weight soon"}, {"id": 784234753, "item_id": 49014672, "url": "https://news.ycombinator.com/item?id=49014672", "by": "InTheArena", "date": "2026-07-22", "stance": "recommend", "quote": "Qwen 27b is a great model", "excerpt": "I have a DGX and a Ryzen AI Max 395 - while I love both of them, there are a few critical things that leave the DGX in use, while the Ryzen \"just\" is my primary homelab server. The biggest thing is prefil numbers, and the performance impact of higher context sizes. Qwen 27b is a great model, nemotron is decent, gemma is workable. But all of them need reasonable context for reasonable outputs. Unfortunitly, as others ", "context": "owns a DGX and Ryzen AI Max 395", "use_case": "local LLM", "switched_from": "", "story": "Nvidia DGX Spark as a daily driver"}]}, {"label": "Tool preference", "summary": "It tends to use bash commands instead of OpenCode's built-in tools unless steered.", "quotes": [{"id": 779633120, "item_id": 48727070, "url": "https://news.ycombinator.com/item?id=48727070", "by": "cpburns2009", "date": "2026-06-29", "stance": "recommend", "quote": "Qwen 27b works well enough in OpenCode though which is what I use.", "excerpt": "If Qwen is finetuned for a hardness, it'll be Qwen Code. Qwen 27b works well enough in OpenCode though which is what I use. My one complaint is it likes to get cute with bash commands instead of OpenCode's built-in tools. I use a skill to steer that.", "context": "", "use_case": "local coding agent", "switched_from": "", "story": "Qwen 3.6 27B is the sweet spot for local development"}]}, {"label": "Tricky distillation", "summary": "Building workflows to train specialists on it is tricky.", "quotes": [{"id": 786069953, "item_id": 49129372, "url": "https://news.ycombinator.com/item?id=49129372", "by": "teravor", "date": "2026-07-31", "stance": "recommend", "quote": "we used to distill GLM 5.2 into Qwen 27B specialists with great success.", "excerpt": "distilling K3 into DS4 flash will likely only be a good idea for specialists. the difference in model capacity is otherwise too large. we used to distill GLM 5.2 into Qwen 27B specialists with great success. developing the workflows is tricky though. we had the advantage of a straightforward mapping function in mind: English -> SQL which was relatively easy to pipeline training for.", "context": "team training specialist models", "use_case": "English to SQL specialist model", "switched_from": "", "story": "DeepSeek-V4-Flash Update"}]}], "who": [{"by": "teravor", "context": "team training specialist models", "url": "https://news.ycombinator.com/item?id=49129372", "date": "2026-07-31"}, {"by": "InTheArena", "context": "owns a DGX and Ryzen AI Max 395", "url": "https://news.ycombinator.com/item?id=49014672", "date": "2026-07-22"}, {"by": "asimovDev", "context": "ran 8-bit MLX version locally", "url": "https://news.ycombinator.com/item?id=48582361", "date": "2026-06-18"}], "tags": ["runs locally", "for agentic coding", "for privacy-sensitive work", "for homelab users", "for training specialist models"], "use_cases": ["coin flip comparison for model decisions", "private local chats", "local model", "local planning", "english to sql specialist model", "local llm"], "compared": [{"slug": "glm", "name": "GLM", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "open-webui", "name": "Open WebUI", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=48860979", "title": "No, Qwen 27B isn't the sweet spot", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=50106023", "title": "Jev meets the Trolley Problem", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49633286", "title": "Culture Becomes a Dark Forest", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49479951", "title": "Run Qwen3.8 27B locally: real numbers from my Mac Studio", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49150470", "title": "Qwen3.8-Max: A New Bar for Coding and Cowork", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49119559", "title": "DeepSeek-V4-Flash Update", "mentions": 1}], "sentiment_by_month": {"2025-11": {}, "2025-12": {}, "2026-01": {}, "2026-02": {}, "2026-03": {}, "2026-04": {}, "2026-05": {}, "2026-06": {"recommend": 2, "neutral": 1}, "2026-07": {"recommend": 5}, "2026-08": {"recommend": 1, "neutral": 1}, "2026-09": {"neutral": 1}, "2026-10": {"neutral": 1}}, "quotes": [{"id": 801750656, "item_id": 50109416, "url": "https://news.ycombinator.com/item?id=50109416", "by": "verdverm", "date": "2026-10-15", "stance": "neutral", "quote": "If little qwen 27b can reply \"50/50, 100% confident\"", "excerpt": "I agree very much, not blindly trusting model output is Rule #2 for interacting with Ai I use the LLM coin flip as a comparison point. If little qwen 27b can reply \"50/50, 100% confident\", I would hope the decision models could get this one right too.", "context": "", "use_case": "coin flip comparison for model decisions", "switched_from": "", "story": "Jev meets the Trolley Problem"}, {"id": 794167776, "item_id": 49635486, "url": "https://news.ycombinator.com/item?id=49635486", "by": "tompark", "date": "2026-09-09", "stance": "neutral", "quote": "except when using local models like Qwen-27B", "excerpt": "Yup. I'm more paranoid about AI training on my chats than most ppl are. While I use frontier models to do research and to build tools, I've tried to avoid discussing details about my projects (e.g. story worldbuilding) except when using local models like Qwen-27B. Of course this slows down progress and has a real opportunity cost, but it would kill all my motivation if I got \"scooped\" like that.", "context": "", "use_case": "private local chats", "switched_from": "", "story": "Culture Becomes a Dark Forest"}, {"id": 791703076, "item_id": 49481442, "url": "https://news.ycombinator.com/item?id=49481442", "by": "wolvoleo", "date": "2026-08-28", "stance": "neutral", "quote": "And I can't run qwen 27b reliably at the moment, I need a better gpu.", "excerpt": "Similar here, except I used Opus 4.8. I don't mind using cloud models to make code as long as I don't put actual private data into the model. And I can't run qwen 27b reliably at the moment, I need a better gpu. It was pretty easy mode like this tbh, though you do need to know what you're doing. I use it with llama-server and openwebui. I guess you can get it a lot more click to go with something like LM Studio thoug", "context": "", "use_case": "local model", "switched_from": "", "story": "Run Qwen3.8 27B locally: real numbers from my Mac Studio"}, {"id": 786437664, "item_id": 49152354, "url": "https://news.ycombinator.com/item?id=49152354", "by": "wickedsight", "date": "2026-08-03", "stance": "recommend", "quote": "I use 27B in plan mode and 35B MoE in act mode.", "excerpt": "I use 27B in plan mode and 35B MoE in act mode. I noticed that is the best balance for me for consistent tool calls and intelligent planning. Takes some time to switch, but it's worth it for me.", "context": "", "use_case": "local planning", "switched_from": "", "story": "Qwen3.8-Max: A New Bar for Coding and Cowork"}, {"id": 786069953, "item_id": 49129372, "url": "https://news.ycombinator.com/item?id=49129372", "by": "teravor", "date": "2026-07-31", "stance": "recommend", "quote": "we used to distill GLM 5.2 into Qwen 27B specialists with great success.", "excerpt": "distilling K3 into DS4 flash will likely only be a good idea for specialists. the difference in model capacity is otherwise too large. we used to distill GLM 5.2 into Qwen 27B specialists with great success. developing the workflows is tricky though. we had the advantage of a straightforward mapping function in mind: English -> SQL which was relatively easy to pipeline training for.", "context": "team training specialist models", "use_case": "English to SQL specialist model", "switched_from": "", "story": "DeepSeek-V4-Flash Update"}, {"id": 784234753, "item_id": 49014672, "url": "https://news.ycombinator.com/item?id=49014672", "by": "InTheArena", "date": "2026-07-22", "stance": "recommend", "quote": "Qwen 27b is a great model", "excerpt": "I have a DGX and a Ryzen AI Max 395 - while I love both of them, there are a few critical things that leave the DGX in use, while the Ryzen \"just\" is my primary homelab server. The biggest thing is prefil numbers, and the performance impact of higher context sizes. Qwen 27b is a great model, nemotron is decent, gemma is workable. But all of them need reasonable context for reasonable outputs. Unfortunitly, as others ", "context": "owns a DGX and Ryzen AI Max 395", "use_case": "local LLM", "switched_from": "", "story": "Nvidia DGX Spark as a daily driver"}, {"id": 783488529, "item_id": 48968033, "url": "https://news.ycombinator.com/item?id=48968033", "by": "eurekin", "date": "2026-07-19", "stance": "recommend", "quote": "27b is a great model, if grounded.", "excerpt": "It's a mcp, so connects quite easily to agents. With mcpo, I also connected it to open-webui (which has better support for OpenAI style tools/functions). Used it in claude code with that mcp plugin set-up too. Only ever used it for managing homelab information, but it met initial expectations. 27b is a great model, if grounded. The query about physical hosts and routing... I haven't found a single hallucination (alto", "context": "", "use_case": "homelab information queries", "switched_from": "", "story": "Qwen3.8 is launching and going open-weight soon"}, {"id": 781786080, "item_id": 48861630, "url": "https://news.ycombinator.com/item?id=48861630", "by": "edyos", "date": "2026-07-10", "stance": "recommend", "quote": "I use this 27B for everyday tasks that aren\u2019t actually coding.", "excerpt": "You\u2019re right about the coding, but not about the rest. I use this 27B for everyday tasks that aren\u2019t actually coding. When it comes to coding, GLM5.2 is, in my view, a marvel; it\u2019s inexpensive and works very well when used properly \u2013 but for everything else, from sorting emails to pre-generation and analysis of varying quality, it does the job that GPT and Claude can\u2019t do for free, and sometimes not quite as well. I ", "context": "", "use_case": "sorting emails, analysis", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}, {"id": 781782097, "item_id": 48861381, "url": "https://news.ycombinator.com/item?id=48861381", "by": "verdverm", "date": "2026-07-10", "stance": "recommend", "quote": "I use this model for agentic work", "excerpt": "Unconvincing, I use this model for agentic work, and while it's obviously not as good, this person is using their personal setup and experience as a justification. 1. Mac is slow at prompt processing, the Spark is much better for coding. Most of the time in coding is spent in prompt processing. 2. DFlash addon improves tgen speed 2-3x. (https://huggingface.co/collections/z-lab/dflash) 3. Harness matters, they gave no", "context": "", "use_case": "agentic work", "switched_from": "", "story": "No, Qwen 27B isn't the sweet spot"}, {"id": 779633120, "item_id": 48727070, "url": "https://news.ycombinator.com/item?id=48727070", "by": "cpburns2009", "date": "2026-06-29", "stance": "recommend", "quote": "Qwen 27b works well enough in OpenCode though which is what I use.", "excerpt": "If Qwen is finetuned for a hardness, it'll be Qwen Code. Qwen 27b works well enough in OpenCode though which is what I use. My one complaint is it likes to get cute with bash commands instead of OpenCode's built-in tools. I use a skill to steer that.", "context": "", "use_case": "local coding agent", "switched_from": "", "story": "Qwen 3.6 27B is the sweet spot for local development"}, {"id": 777317776, "item_id": 48582361, "url": "https://news.ycombinator.com/item?id=48582361", "by": "asimovDev", "date": "2026-06-18", "stance": "neutral", "quote": "It succesfully identified how it worked even the obfuscated classes and methods.", "excerpt": "I used Qwen 27b 8 bit MLX version on a decompiled android APK recently. It succesfully identified how it worked even the obfuscated classes and methods. It wrote a 1000 line documentation with examples but the time was dreadful. At some point it slowed down to 5 t/s so the whole thing took over an hour , the writing of documentation alone was over 40 minutes, fans blasting the entire time.", "context": "ran 8-bit MLX version locally", "use_case": "analyzing decompiled Android APK", "switched_from": "", "story": "Local Qwen isn't a worse Opus, it's a different tool"}, {"id": 776968864, "item_id": 48560554, "url": "https://news.ycombinator.com/item?id=48560554", "by": "freehorse", "date": "2026-06-16", "stance": "recommend", "quote": "I have been using the qwen 27b and it is great", "excerpt": "> You have MoE models (gemma 26b, qwen 35b, north mini code 30b) who are pretty fast, but make a lot of mistakes This is sadly also my experience. I wish we had some MoE models with a higher ratio of active parameters per total. My experience is that the newer MoE models that can run in a 64b laptop have too few active parameters to be useful outside narrower, specific tasks. Mixtral 8x7b was a 14b active parameter (", "context": "", "use_case": "local inference on laptop", "switched_from": "", "story": "Running local models is good now"}]}