{"slug": "qwen3", "name": "Qwen3", "category": "ai", "category_name": "AI models and assistants", "description": "Alibaba's Qwen3 family of open-weight large language models, in dense and mixture-of-experts variants.", "recommenders": 7, "commenters": 9, "warners": 0, "score": 6.209, "rising": -0.96, "first": "2025-11-22", "latest": "2026-09-10", "why": "Liked as open-weights models you can run at home, from small quantized builds to promising MoE variants, and as robust agents with tools like opencode.", "why_is_quote": false, "has_verdict": true, "sentiment": {"recommend": 7, "neutral": 3, "compare": 0, "warn": 0}, "spark": [1, 0, 0, 1, 1, 1, 0, 2, 0, 0, 1, 0], "listed": true, "loved": ["Runs locally", "Small model options"], "caveat": "Slower locally", "short_quote": "I tend to rely on Qwen3 (fast) for quick questions and Qwen3 (reasoning) when I want a more structured approach", "icon": true, "domain": null, "github_repo": null, "official_url": "https://github.com/QwenLM/Qwen3", "pricing": "open source", "aliases": ["Qwen3", "Qwen3 1.7B", "Qwen3-235B", "Qwen3 235B"], "verdict": [{"text": "Liked as open-weights models you can run at home", "cites": [{"id": 736217890, "item_id": 46013618, "url": "https://news.ycombinator.com/item?id=46013618", "by": "Roark66", "date": "2025-11-22", "stance": "recommend", "quote": "open weights models like Qwen3 and kimik2 exceeded their capability and you can run them at home", "excerpt": "I'm especially annoyed that this is most likely intentional. (Not at all)openAI saw they are getting behind their competitors (gpt 5 and 5.1 were progressively worse for my use case - actual problem solving and tweaking existing scripts) are getting better. (Claude and sonnet were miles ahead and I used gpt only due to lower price). Now not only open weights models like Qwen3 and kimik2 exceeded their capability and ", "context": "", "use_case": "local models", "switched_from": "", "story": "Samsung's 60% DRAM price hike signals a new phase of global memory tightening"}, {"id": 766643088, "item_id": 47915193, "url": "https://news.ycombinator.com/item?id=47915193", "by": "cyanydeez", "date": "2026-04-26", "stance": "recommend", "quote": "QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins", "excerpt": "unsurprising. QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins, prompts etc. It might take longer, but again, people who've suckled on these larger models should really step down (many of them will be forced too with enshittification comes) and take a look at where these models can replace cloud providers.", "context": "", "use_case": "local LLM coding agent", "switched_from": "", "story": "The New Linux Kernel AI Bot Uncovering Bugs Is a Local LLM on Framework Desktop"}]}, {"text": ", from small quantized builds to promising MoE variants,", "cites": [{"id": 779360512, "item_id": 48710032, "url": "https://news.ycombinator.com/item?id=48710032", "by": "macwhisperer", "date": "2026-06-28", "stance": "recommend", "quote": "qwen3 1.7b- q4_k_m is your best best for that size", "excerpt": "qwen3 1.7b- q4_k_m is your best best for that size", "context": "", "use_case": "small local LLM", "switched_from": "", "story": "Ask HN: Best local LLM under 2B paramater and consuming RAM less than 3gb"}, {"id": 776948384, "item_id": 48559274, "url": "https://news.ycombinator.com/item?id=48559274", "by": "ryukoposting", "date": "2026-06-16", "stance": "neutral", "quote": "the heavily quantized Qwen3 models I can cram onto my 3060 Ti", "excerpt": "I found that, with the heavily quantized Qwen3 models I can cram onto my 3060 Ti, telling the model to use its tools in the system prompt made it a lot more likely to actually do it. YMMV of course, but give it a shot.", "context": "", "use_case": "local inference with tools", "switched_from": "", "story": "Running local models is good now"}, {"id": 794202775, "item_id": 49637673, "url": "https://news.ycombinator.com/item?id=49637673", "by": "softwarewright", "date": "2026-09-10", "stance": "recommend", "quote": "recent qwen3.x MoE models look very promising", "excerpt": "Claude Code, Codex, opencode (Z.ai GLM-5.x), sometimes Gemini -- most of my work is at Mac iTerm2 CLI or Arch Linux terminal. I sometimes use emacs as the coding agent, but rarely JetBrains IDE or vscode (more for reading/debugging code). I supplement this with my own vibe-coded tools that help agents plan, perform sagas/steps, stay on track, tools that check the code produced, tools that check the produced documenta", "context": "", "use_case": "local coding model", "switched_from": "", "story": "Ask HN: What's your AI coding set up?"}, {"id": 779429584, "item_id": 48714349, "url": "https://news.ycombinator.com/item?id=48714349", "by": "adithyaharish", "date": "2026-06-28", "stance": "recommend", "quote": "just checked it out, has mixture of experts, mit fit my use case!", "excerpt": "just checked it out, has mixture of experts, mit fit my use case!", "context": "", "use_case": "small local LLM", "switched_from": "", "story": "Ask HN: Best local LLM under 2B paramater and consuming RAM less than 3gb"}]}, {"text": " and as robust agents with tools like opencode.", "cites": [{"id": 766643088, "item_id": 47915193, "url": "https://news.ycombinator.com/item?id=47915193", "by": "cyanydeez", "date": "2026-04-26", "stance": "recommend", "quote": "QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins", "excerpt": "unsurprising. QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins, prompts etc. It might take longer, but again, people who've suckled on these larger models should really step down (many of them will be forced too with enshittification comes) and take a look at where these models can replace cloud providers.", "context": "", "use_case": "local LLM coding agent", "switched_from": "", "story": "The New Linux Kernel AI Bot Uncovering Bugs Is a Local LLM on Framework Desktop"}]}, {"text": " No single complaint dominates: one user says local use can take longer than big cloud models,", "cites": [{"id": 766643088, "item_id": 47915193, "url": "https://news.ycombinator.com/item?id=47915193", "by": "cyanydeez", "date": "2026-04-26", "stance": "recommend", "quote": "QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins", "excerpt": "unsurprising. QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins, prompts etc. It might take longer, but again, people who've suckled on these larger models should really step down (many of them will be forced too with enshittification comes) and take a look at where these models can replace cloud providers.", "context": "", "use_case": "local LLM coding agent", "switched_from": "", "story": "The New Linux Kernel AI Bot Uncovering Bugs Is a Local LLM on Framework Desktop"}]}, {"text": " another prefers Qwen 2.5 Coder for executing code, and heavily quantized versions need explicit prompting to use tools.", "cites": [{"id": 741651078, "item_id": 46353192, "url": "https://news.ycombinator.com/item?id=46353192", "by": "Roark66", "date": "2025-12-22", "stance": "neutral", "quote": "Qwen3-235B - between Qwen 2.5 and Sonet", "excerpt": "I found the winning combination is to use all of them in this way: - first you need a vendor agnostic tool like opencode (I had to add my own vendors as it didn't support it out of the box properly) - second you set up agents with different models. I use: - for architecture and planning - opus, Sonet, gpt 5.2, gemini3 (depending on specifics, for example I found got better in troubleshooting, Sonet better in pure cod", "context": "", "use_case": "executing plans", "switched_from": "", "story": "A guide to local coding models"}, {"id": 776948384, "item_id": 48559274, "url": "https://news.ycombinator.com/item?id=48559274", "by": "ryukoposting", "date": "2026-06-16", "stance": "neutral", "quote": "the heavily quantized Qwen3 models I can cram onto my 3060 Ti", "excerpt": "I found that, with the heavily quantized Qwen3 models I can cram onto my 3060 Ti, telling the model to use its tools in the system prompt made it a lot more likely to actually do it. YMMV of course, but give it a shot.", "context": "", "use_case": "local inference with tools", "switched_from": "", "story": "Running local models is good now"}]}], "praise": [{"label": "Runs locally", "summary": "Open weights let users run it at home or cheaply through providers instead of relying on cloud models.", "quotes": [{"id": 736217890, "item_id": 46013618, "url": "https://news.ycombinator.com/item?id=46013618", "by": "Roark66", "date": "2025-11-22", "stance": "recommend", "quote": "open weights models like Qwen3 and kimik2 exceeded their capability and you can run them at home", "excerpt": "I'm especially annoyed that this is most likely intentional. (Not at all)openAI saw they are getting behind their competitors (gpt 5 and 5.1 were progressively worse for my use case - actual problem solving and tweaking existing scripts) are getting better. (Claude and sonnet were miles ahead and I used gpt only due to lower price). Now not only open weights models like Qwen3 and kimik2 exceeded their capability and ", "context": "", "use_case": "local models", "switched_from": "", "story": "Samsung's 60% DRAM price hike signals a new phase of global memory tightening"}, {"id": 766643088, "item_id": 47915193, "url": "https://news.ycombinator.com/item?id=47915193", "by": "cyanydeez", "date": "2026-04-26", "stance": "recommend", "quote": "QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins", "excerpt": "unsurprising. QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins, prompts etc. It might take longer, but again, people who've suckled on these larger models should really step down (many of them will be forced too with enshittification comes) and take a look at where these models can replace cloud providers.", "context": "", "use_case": "local LLM coding agent", "switched_from": "", "story": "The New Linux Kernel AI Bot Uncovering Bugs Is a Local LLM on Framework Desktop"}, {"id": 776948384, "item_id": 48559274, "url": "https://news.ycombinator.com/item?id=48559274", "by": "ryukoposting", "date": "2026-06-16", "stance": "neutral", "quote": "the heavily quantized Qwen3 models I can cram onto my 3060 Ti", "excerpt": "I found that, with the heavily quantized Qwen3 models I can cram onto my 3060 Ti, telling the model to use its tools in the system prompt made it a lot more likely to actually do it. YMMV of course, but give it a shot.", "context": "", "use_case": "local inference with tools", "switched_from": "", "story": "Running local models is good now"}]}, {"label": "Small model options", "summary": "Small and heavily quantized versions fit modest hardware like a 3060 Ti.", "quotes": [{"id": 779360512, "item_id": 48710032, "url": "https://news.ycombinator.com/item?id=48710032", "by": "macwhisperer", "date": "2026-06-28", "stance": "recommend", "quote": "qwen3 1.7b- q4_k_m is your best best for that size", "excerpt": "qwen3 1.7b- q4_k_m is your best best for that size", "context": "", "use_case": "small local LLM", "switched_from": "", "story": "Ask HN: Best local LLM under 2B paramater and consuming RAM less than 3gb"}, {"id": 776948384, "item_id": 48559274, "url": "https://news.ycombinator.com/item?id=48559274", "by": "ryukoposting", "date": "2026-06-16", "stance": "neutral", "quote": "the heavily quantized Qwen3 models I can cram onto my 3060 Ti", "excerpt": "I found that, with the heavily quantized Qwen3 models I can cram onto my 3060 Ti, telling the model to use its tools in the system prompt made it a lot more likely to actually do it. YMMV of course, but give it a shot.", "context": "", "use_case": "local inference with tools", "switched_from": "", "story": "Running local models is good now"}]}, {"label": "MoE models", "summary": "The mixture-of-experts variants are described as promising and a good fit.", "quotes": [{"id": 794202775, "item_id": 49637673, "url": "https://news.ycombinator.com/item?id=49637673", "by": "softwarewright", "date": "2026-09-10", "stance": "recommend", "quote": "recent qwen3.x MoE models look very promising", "excerpt": "Claude Code, Codex, opencode (Z.ai GLM-5.x), sometimes Gemini -- most of my work is at Mac iTerm2 CLI or Arch Linux terminal. I sometimes use emacs as the coding agent, but rarely JetBrains IDE or vscode (more for reading/debugging code). I supplement this with my own vibe-coded tools that help agents plan, perform sagas/steps, stay on track, tools that check the code produced, tools that check the produced documenta", "context": "", "use_case": "local coding model", "switched_from": "", "story": "Ask HN: What's your AI coding set up?"}, {"id": 779429584, "item_id": 48714349, "url": "https://news.ycombinator.com/item?id=48714349", "by": "adithyaharish", "date": "2026-06-28", "stance": "recommend", "quote": "just checked it out, has mixture of experts, mit fit my use case!", "excerpt": "just checked it out, has mixture of experts, mit fit my use case!", "context": "", "use_case": "small local LLM", "switched_from": "", "story": "Ask HN: Best local LLM under 2B paramater and consuming RAM less than 3gb"}]}, {"label": "Agent robustness", "summary": "Seen as robust when driven by opencode with bash scripts, plugins and prompts.", "quotes": [{"id": 766643088, "item_id": 47915193, "url": "https://news.ycombinator.com/item?id=47915193", "by": "cyanydeez", "date": "2026-04-26", "stance": "recommend", "quote": "QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins", "excerpt": "unsurprising. QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins, prompts etc. It might take longer, but again, people who've suckled on these larger models should really step down (many of them will be forced too with enshittification comes) and take a look at where these models can replace cloud providers.", "context": "", "use_case": "local LLM coding agent", "switched_from": "", "story": "The New Linux Kernel AI Bot Uncovering Bugs Is a Local LLM on Framework Desktop"}]}, {"label": "Fast and reasoning", "summary": "One user relies on the fast mode for quick questions and the reasoning mode for structured research.", "quotes": [{"id": 755675056, "item_id": 47229691, "url": "https://news.ycombinator.com/item?id=47229691", "by": "mongrelion", "date": "2026-03-03", "stance": "recommend", "quote": "I tend to rely on Qwen3 (fast) for quick questions and Qwen3 (reasoning) when I want a more structured approach", "excerpt": "Through my Kagi subscription I get access to quite a few models [1] but I tend to rely on Qwen3 (fast) for quick questions and Qwen3 (reasoning) when I want a more structured approach, for example, when I am researching a topic. I have tried the same approach with Kimi K2.5 and GLM 5 but I keep going back fo Qwen3. I also have access to Perplexity which is quite decent to be honest, but I prefer to keep everything in", "context": "Kagi subscriber", "use_case": "quick questions and research", "switched_from": "", "story": "Ask HN: What Online LLM / Chat do you use?"}]}, {"label": "Upgrade over 2.5", "summary": "Widely regarded as a significant upgrade over Qwen2.5.", "quotes": [{"id": 753436480, "item_id": 47089780, "url": "https://news.ycombinator.com/item?id=47089780", "by": "magicalhippo", "date": "2026-02-20", "stance": "recommend", "quote": "Yet Qwen3 was widely regarded as a significant upgrade to Qwen2.5.", "excerpt": "From what I've gathered, they've been mostly training limited. Better training methods and cleaner training data allows smaller models to rival or outperform larger models training with older methods and lower-quality training data. For example, the Qwen3 technical report[1] says that the Qwen3 models are architecturally very similar to Qwen2.5, with the main change being a tweak in the attention layers to stabilize ", "context": "", "use_case": "example of training-driven gains", "switched_from": "", "story": "Consistency diffusion language models: Up to 14x faster, no quality loss"}]}], "caveats": [{"label": "Slower locally", "summary": "Running these models instead of larger cloud models might take longer.", "quotes": [{"id": 766643088, "item_id": 47915193, "url": "https://news.ycombinator.com/item?id=47915193", "by": "cyanydeez", "date": "2026-04-26", "stance": "recommend", "quote": "QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins", "excerpt": "unsurprising. QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins, prompts etc. It might take longer, but again, people who've suckled on these larger models should really step down (many of them will be forced too with enshittification comes) and take a look at where these models can replace cloud providers.", "context": "", "use_case": "local LLM coding agent", "switched_from": "", "story": "The New Linux Kernel AI Bot Uncovering Bugs Is a Local LLM on Framework Desktop"}]}, {"label": "Coding execution", "summary": "One user finds Qwen 2.5 Coder 30B better than Qwen3 for executing coding plans.", "quotes": [{"id": 741651078, "item_id": 46353192, "url": "https://news.ycombinator.com/item?id=46353192", "by": "Roark66", "date": "2025-12-22", "stance": "neutral", "quote": "Qwen3-235B - between Qwen 2.5 and Sonet", "excerpt": "I found the winning combination is to use all of them in this way: - first you need a vendor agnostic tool like opencode (I had to add my own vendors as it didn't support it out of the box properly) - second you set up agents with different models. I use: - for architecture and planning - opus, Sonet, gpt 5.2, gemini3 (depending on specifics, for example I found got better in troubleshooting, Sonet better in pure cod", "context": "", "use_case": "executing plans", "switched_from": "", "story": "A guide to local coding models"}]}, {"label": "Tool use prompting", "summary": "Heavily quantized versions were more reliable at using tools when the system prompt told them to.", "quotes": [{"id": 776948384, "item_id": 48559274, "url": "https://news.ycombinator.com/item?id=48559274", "by": "ryukoposting", "date": "2026-06-16", "stance": "neutral", "quote": "the heavily quantized Qwen3 models I can cram onto my 3060 Ti", "excerpt": "I found that, with the heavily quantized Qwen3 models I can cram onto my 3060 Ti, telling the model to use its tools in the system prompt made it a lot more likely to actually do it. YMMV of course, but give it a shot.", "context": "", "use_case": "local inference with tools", "switched_from": "", "story": "Running local models is good now"}]}], "who": [{"by": "micstradev", "context": "builder of the project", "url": "https://news.ycombinator.com/item?id=48469974", "date": "2026-06-09"}, {"by": "mongrelion", "context": "Kagi subscriber", "url": "https://news.ycombinator.com/item?id=47229691", "date": "2026-03-03"}], "tags": ["for local inference", "for coding agents", "for low-VRAM GPUs", "for research questions"], "use_cases": ["small local llm", "local coding model", "local inference with tools", "free tier model", "local llm coding agent", "quick questions and research"], "compared": [{"slug": "opencode", "name": "OpenCode", "co": 3, "switched_from": 0, "switched_to": 0}, {"slug": "gemini", "name": "Gemini", "co": 3, "switched_from": 0, "switched_to": 0}, {"slug": "gpt-5", "name": "GPT-5", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "claude-sonnet", "name": "Claude Sonnet", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "kimi", "name": "Kimi", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=48708262", "title": "Ask HN: Best local LLM under 2B paramater and consuming RAM less than 3gb", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=49636239", "title": "Ask HN: What's your AI coding set up?", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48555993", "title": "Running local models is good now", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48459913", "title": "Show HN: Learn from 30 historical figures, open source, nonprofit, self-hosted", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=47914388", "title": "The New Linux Kernel AI Bot Uncovering Bugs Is a Local LLM on Framework Desktop", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=47227046", "title": "Ask HN: What Online LLM / Chat do you use?", "mentions": 1}], "sentiment_by_month": {"2025-11": {"recommend": 1}, "2025-12": {"neutral": 1}, "2026-01": {}, "2026-02": {"recommend": 1}, "2026-03": {"recommend": 1}, "2026-04": {"recommend": 1}, "2026-05": {}, "2026-06": {"neutral": 2, "recommend": 2}, "2026-07": {}, "2026-08": {}, "2026-09": {"recommend": 1}, "2026-10": {}}, "quotes": [{"id": 794202775, "item_id": 49637673, "url": "https://news.ycombinator.com/item?id=49637673", "by": "softwarewright", "date": "2026-09-10", "stance": "recommend", "quote": "recent qwen3.x MoE models look very promising", "excerpt": "Claude Code, Codex, opencode (Z.ai GLM-5.x), sometimes Gemini -- most of my work is at Mac iTerm2 CLI or Arch Linux terminal. I sometimes use emacs as the coding agent, but rarely JetBrains IDE or vscode (more for reading/debugging code). I supplement this with my own vibe-coded tools that help agents plan, perform sagas/steps, stay on track, tools that check the code produced, tools that check the produced documenta", "context": "", "use_case": "local coding model", "switched_from": "", "story": "Ask HN: What's your AI coding set up?"}, {"id": 779429584, "item_id": 48714349, "url": "https://news.ycombinator.com/item?id=48714349", "by": "adithyaharish", "date": "2026-06-28", "stance": "recommend", "quote": "just checked it out, has mixture of experts, mit fit my use case!", "excerpt": "just checked it out, has mixture of experts, mit fit my use case!", "context": "", "use_case": "small local LLM", "switched_from": "", "story": "Ask HN: Best local LLM under 2B paramater and consuming RAM less than 3gb"}, {"id": 779360512, "item_id": 48710032, "url": "https://news.ycombinator.com/item?id=48710032", "by": "macwhisperer", "date": "2026-06-28", "stance": "recommend", "quote": "qwen3 1.7b- q4_k_m is your best best for that size", "excerpt": "qwen3 1.7b- q4_k_m is your best best for that size", "context": "", "use_case": "small local LLM", "switched_from": "", "story": "Ask HN: Best local LLM under 2B paramater and consuming RAM less than 3gb"}, {"id": 776948384, "item_id": 48559274, "url": "https://news.ycombinator.com/item?id=48559274", "by": "ryukoposting", "date": "2026-06-16", "stance": "neutral", "quote": "the heavily quantized Qwen3 models I can cram onto my 3060 Ti", "excerpt": "I found that, with the heavily quantized Qwen3 models I can cram onto my 3060 Ti, telling the model to use its tools in the system prompt made it a lot more likely to actually do it. YMMV of course, but give it a shot.", "context": "", "use_case": "local inference with tools", "switched_from": "", "story": "Running local models is good now"}, {"id": 775519585, "item_id": 48469974, "url": "https://news.ycombinator.com/item?id=48469974", "by": "micstradev", "date": "2026-06-09", "stance": "neutral", "quote": "It is the base model, Qwen3 235B for the free tier", "excerpt": "We have no custom models and no fine tuning. It is the base model, Qwen3 235B for the free tier and we recommend Qwen3.6 27B for the local mode. So the figures are data, not weights. We have instruction files for every figure and additional voice profiles for the councils. The work was put into the iterations to improve the instructions. It is possible to fetch them from the CDN, for example: https://media.agoracosmi", "context": "builder of the project", "use_case": "free tier model", "switched_from": "", "story": "Show HN: Learn from 30 historical figures, open source, nonprofit, self-hosted"}, {"id": 766643088, "item_id": 47915193, "url": "https://news.ycombinator.com/item?id=47915193", "by": "cyanydeez", "date": "2026-04-26", "stance": "recommend", "quote": "QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins", "excerpt": "unsurprising. QWEN3* models appear to be very robust and can be prompted with opencode using basic bash scripts and robust plugins, prompts etc. It might take longer, but again, people who've suckled on these larger models should really step down (many of them will be forced too with enshittification comes) and take a look at where these models can replace cloud providers.", "context": "", "use_case": "local LLM coding agent", "switched_from": "", "story": "The New Linux Kernel AI Bot Uncovering Bugs Is a Local LLM on Framework Desktop"}, {"id": 755675056, "item_id": 47229691, "url": "https://news.ycombinator.com/item?id=47229691", "by": "mongrelion", "date": "2026-03-03", "stance": "recommend", "quote": "I tend to rely on Qwen3 (fast) for quick questions and Qwen3 (reasoning) when I want a more structured approach", "excerpt": "Through my Kagi subscription I get access to quite a few models [1] but I tend to rely on Qwen3 (fast) for quick questions and Qwen3 (reasoning) when I want a more structured approach, for example, when I am researching a topic. I have tried the same approach with Kimi K2.5 and GLM 5 but I keep going back fo Qwen3. I also have access to Perplexity which is quite decent to be honest, but I prefer to keep everything in", "context": "Kagi subscriber", "use_case": "quick questions and research", "switched_from": "", "story": "Ask HN: What Online LLM / Chat do you use?"}, {"id": 753436480, "item_id": 47089780, "url": "https://news.ycombinator.com/item?id=47089780", "by": "magicalhippo", "date": "2026-02-20", "stance": "recommend", "quote": "Yet Qwen3 was widely regarded as a significant upgrade to Qwen2.5.", "excerpt": "From what I've gathered, they've been mostly training limited. Better training methods and cleaner training data allows smaller models to rival or outperform larger models training with older methods and lower-quality training data. For example, the Qwen3 technical report[1] says that the Qwen3 models are architecturally very similar to Qwen2.5, with the main change being a tweak in the attention layers to stabilize ", "context": "", "use_case": "example of training-driven gains", "switched_from": "", "story": "Consistency diffusion language models: Up to 14x faster, no quality loss"}, {"id": 741651078, "item_id": 46353192, "url": "https://news.ycombinator.com/item?id=46353192", "by": "Roark66", "date": "2025-12-22", "stance": "neutral", "quote": "Qwen3-235B - between Qwen 2.5 and Sonet", "excerpt": "I found the winning combination is to use all of them in this way: - first you need a vendor agnostic tool like opencode (I had to add my own vendors as it didn't support it out of the box properly) - second you set up agents with different models. I use: - for architecture and planning - opus, Sonet, gpt 5.2, gemini3 (depending on specifics, for example I found got better in troubleshooting, Sonet better in pure cod", "context": "", "use_case": "executing plans", "switched_from": "", "story": "A guide to local coding models"}, {"id": 736217890, "item_id": 46013618, "url": "https://news.ycombinator.com/item?id=46013618", "by": "Roark66", "date": "2025-11-22", "stance": "recommend", "quote": "open weights models like Qwen3 and kimik2 exceeded their capability and you can run them at home", "excerpt": "I'm especially annoyed that this is most likely intentional. (Not at all)openAI saw they are getting behind their competitors (gpt 5 and 5.1 were progressively worse for my use case - actual problem solving and tweaking existing scripts) are getting better. (Claude and sonnet were miles ahead and I used gpt only due to lower price). Now not only open weights models like Qwen3 and kimik2 exceeded their capability and ", "context": "", "use_case": "local models", "switched_from": "", "story": "Samsung's 60% DRAM price hike signals a new phase of global memory tightening"}]}