{"slug": "mlx", "name": "MLX", "category": "dev-libs", "category_name": "Libraries and frameworks", "description": "", "recommenders": 4, "commenters": 14, "warners": 3, "score": 2.919, "rising": 4.0, "first": "2026-01-15", "latest": "2026-10-28", "why": "I tried to tune it to use MLX on Mac GPUs, and it's so much faster.", "why_is_quote": true, "has_verdict": false, "sentiment": {"recommend": 4, "neutral": 7, "compare": 1, "warn": 3}, "spark": [0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 1], "listed": true, "loved": [], "caveat": "", "short_quote": "I tried to tune it to use MLX on Mac GPUs, and it's so much faster.", "icon": true, "domain": null, "github_repo": "github.com/ml-explore/mlx", "official_url": "https://github.com/ml-explore/mlx", "pricing": "", "aliases": ["MLX"], "verdict": null, "praise": [{"label": "", "summary": "", "quotes": [{"id": 804509793, "item_id": 50281862, "url": "https://news.ycombinator.com/item?id=50281862", "by": "inom155", "date": "2026-10-28", "stance": "recommend", "quote": "I tried to tune it to use MLX on Mac GPUs, and it's so much faster.", "excerpt": "Honestly, I was very impressed with this. I did find the latency was quite high, so I tried to tune it to use MLX on Mac GPUs, and it's so much faster. Very cool either way, we're getting close to a full AI interactive tool that's offline and private! https://github.com/skorotkiewicz/lfm3-chat/pull/1", "context": "", "use_case": "faster inference on Mac", "switched_from": "", "story": "Show HN: LFM3 Video Chat"}, {"id": 796356993, "item_id": 49772312, "url": "https://news.ycombinator.com/item?id=49772312", "by": "aetherspawn", "date": "2026-09-20", "stance": "recommend", "quote": "I use LM Studio with MLX support, which is specifically faster on M series Macs.", "excerpt": "Yes, I use LM Studio with MLX support, which is specifically faster on M series Macs. I am not sure if llama is the same, but I guess what I\u2019m saying is if you want the performance to be good on M series you have to use models packaged in the right format.", "context": "M series Mac user", "use_case": "fast inference on Apple Silicon", "switched_from": "", "story": "Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint"}, {"id": 791011600, "item_id": 49438225, "url": "https://news.ycombinator.com/item?id=49438225", "by": "EagnaIonat", "date": "2026-08-25", "stance": "recommend", "quote": "runs inference models fine, especially those using MLX", "excerpt": "Your statement is not really true any more. I have an old M1 Max with 32GB that runs inference models fine, especially those using MLX. Dedicated graphics cards are faster but not to the point that it matters. That\u2019s a six year old machine. My newer machine M5 max 128GB will far outperform your typical 32GB gaming card once the model exceeds memory. To even get close to that on a dedicated graphics card you are payin", "context": "", "use_case": "local model inference on Mac", "switched_from": "", "story": "New Mac mini, featuring M6 and M5 Pro"}, {"id": 788243522, "item_id": 49265220, "url": "https://news.ycombinator.com/item?id=49265220", "by": "mark_l_watson", "date": "2026-08-11", "stance": "recommend", "quote": "runs so well on Apple Silicon using MLX!", "excerpt": "I love the wave of new small model releases. Pleasantly surprising that an NVIDIA model runs so well on Apple Silicon using MLX! I was using nemotron-3.5-lightning:30b-mlx with OpenCode on my old (cheap) Mac this morning and no bad experiences except for running slowly.", "context": "", "use_case": "running models on Apple Silicon", "switched_from": "", "story": "Nvidia Nemotron 3.5 lightning and Nemo Switchyard"}]}], "caveats": [{"label": "", "summary": "", "quotes": [{"id": 788403256, "item_id": 49275203, "url": "https://news.ycombinator.com/item?id=49275203", "by": "ljosifov", "date": "2026-08-12", "stance": "warn", "quote": "mlx on the apple ASI (that never quite work out for me)", "excerpt": "Atm OMP (oh-my-pi) with opencode-go subscription using deepseek-v4-flash-0731 (and mimo-v2.5 as advisor). Used pi before (still good, but too minimalist for all general use; works on Android under Termux btw) and OpenCode (which is ok too). I've used Codex the longest, but the $20 sub doesn't last that long. Before that Claude (and Claude on the web - about Nov'25), but nowadays with the new models not possible to do", "context": "", "use_case": "local inference on Apple silicon", "switched_from": "", "story": "llama.cpp"}, {"id": 775316177, "item_id": 48457261, "url": "https://news.ycombinator.com/item?id=48457261", "by": "wahnfrieden", "date": "2026-06-08", "stance": "warn", "quote": "MLX is not for end user deployment.", "excerpt": "I use CoreML for models designed for other platforms. I port the models to it but it works for that without much trouble. MLX is not for end user deployment.", "context": "", "use_case": "", "switched_from": "", "story": "Apple Core AI Framework"}, {"id": 769439044, "item_id": 48089940, "url": "https://news.ycombinator.com/item?id=48089940", "by": "quacker", "date": "2026-05-10", "stance": "warn", "quote": "I tried 4bit MLX quants of Qwen 3.5 9B but it eventually would crash due to insufficient memory.", "excerpt": "I could have used this article before I spent the weekend arriving to the same conclusion! Same laptop, and my contrived test was having it fix 50 or so lint errors in a small vibe-coded C++ repo. I wanted it to be able to handle a bunch of small tasks without getting stuck too often. GPT OSS 20B was usable but slow, and actually frequently made mistakes like adding or duplicating statements unnecessarily, listing th", "context": "", "use_case": "running local models", "switched_from": "", "story": "Running local models on an M4 with 24GB memory"}]}], "who": [{"by": "aetherspawn", "context": "M series Mac user", "url": "https://news.ycombinator.com/item?id=49772312", "date": "2026-09-20"}, {"by": "__jf__", "context": "runs local RAG", "url": "https://news.ycombinator.com/item?id=46630180", "date": "2026-01-15"}], "tags": [], "use_cases": ["running models on apple silicon", "faster inference on mac", "on-device inference", "running models on apple hardware", "fast inference on apple silicon", "local model inference on mac"], "compared": [{"slug": "llama-cpp", "name": "llama.cpp", "co": 4, "switched_from": 0, "switched_to": 1}, {"slug": "opencode", "name": "OpenCode", "co": 3, "switched_from": 0, "switched_to": 0}, {"slug": "core-ml", "name": "Core ML", "co": 2, "switched_from": 0, "switched_to": 0}, {"slug": "claude-code", "name": "Claude Code", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=50280966", "title": "Show HN: LFM3 Video Chat", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=50069337", "title": "Show HN: Use your OpenAI subscription to create audiobooks", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49774338", "title": "Is Jev the general-purpose classifier we've been waiting for?", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49746618", "title": "Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49433450", "title": "New Mac mini, featuring M6 and M5 Pro", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49267928", "title": "llama.cpp", "mentions": 1}], "sentiment_by_month": {"2025-11": {}, "2025-12": {}, "2026-01": {"neutral": 1}, "2026-02": {"neutral": 1}, "2026-03": {"neutral": 2}, "2026-04": {"neutral": 1}, "2026-05": {"warn": 1}, "2026-06": {"warn": 1, "compare": 1}, "2026-07": {}, "2026-08": {"recommend": 2, "warn": 1}, "2026-09": {"recommend": 1, "neutral": 1}, "2026-10": {"neutral": 1, "recommend": 1}}, "quotes": [{"id": 804509793, "item_id": 50281862, "url": "https://news.ycombinator.com/item?id=50281862", "by": "inom155", "date": "2026-10-28", "stance": "recommend", "quote": "I tried to tune it to use MLX on Mac GPUs, and it's so much faster.", "excerpt": "Honestly, I was very impressed with this. I did find the latency was quite high, so I tried to tune it to use MLX on Mac GPUs, and it's so much faster. Very cool either way, we're getting close to a full AI interactive tool that's offline and private! https://github.com/skorotkiewicz/lfm3-chat/pull/1", "context": "", "use_case": "faster inference on Mac", "switched_from": "", "story": "Show HN: LFM3 Video Chat"}, {"id": 801284242, "item_id": 50080265, "url": "https://news.ycombinator.com/item?id=50080265", "by": "mier", "date": "2026-10-13", "stance": "neutral", "quote": "using MLX and Kokoro", "excerpt": "I use my hobby app PageHum to read short to medium articles in devices using MLX and Kokoro instead of. Give it a try if anyone interested in https://github.com/Mieraidihaimu/pagehum (Xcode 27+)", "context": "", "use_case": "on-device inference", "switched_from": "", "story": "Show HN: Use your OpenAI subscription to create audiobooks"}, {"id": 796389426, "item_id": 49774339, "url": "https://news.ycombinator.com/item?id=49774339", "by": "kncvetko", "date": "2026-09-20", "stance": "neutral", "quote": "running on a consumer Apple laptop with MLX as the platform", "excerpt": "Is Jev the general-purpose classifier we\u2019ve been waiting for? That would be the thing that could create a real dent in the universe: a classifier that can classify without first collecting a massive statistical dataset and training it with whatever algorithm fits best; be it BERT or good old logistic regression. Long story short: no. The more nuanced answer is: it depends. Over the years, we\u2019ve built classification s", "context": "", "use_case": "running models on Apple hardware", "switched_from": "", "story": "Is Jev the general-purpose classifier we've been waiting for?"}, {"id": 796356993, "item_id": 49772312, "url": "https://news.ycombinator.com/item?id=49772312", "by": "aetherspawn", "date": "2026-09-20", "stance": "recommend", "quote": "I use LM Studio with MLX support, which is specifically faster on M series Macs.", "excerpt": "Yes, I use LM Studio with MLX support, which is specifically faster on M series Macs. I am not sure if llama is the same, but I guess what I\u2019m saying is if you want the performance to be good on M series you have to use models packaged in the right format.", "context": "M series Mac user", "use_case": "fast inference on Apple Silicon", "switched_from": "", "story": "Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint"}, {"id": 791011600, "item_id": 49438225, "url": "https://news.ycombinator.com/item?id=49438225", "by": "EagnaIonat", "date": "2026-08-25", "stance": "recommend", "quote": "runs inference models fine, especially those using MLX", "excerpt": "Your statement is not really true any more. I have an old M1 Max with 32GB that runs inference models fine, especially those using MLX. Dedicated graphics cards are faster but not to the point that it matters. That\u2019s a six year old machine. My newer machine M5 max 128GB will far outperform your typical 32GB gaming card once the model exceeds memory. To even get close to that on a dedicated graphics card you are payin", "context": "", "use_case": "local model inference on Mac", "switched_from": "", "story": "New Mac mini, featuring M6 and M5 Pro"}, {"id": 788403256, "item_id": 49275203, "url": "https://news.ycombinator.com/item?id=49275203", "by": "ljosifov", "date": "2026-08-12", "stance": "warn", "quote": "mlx on the apple ASI (that never quite work out for me)", "excerpt": "Atm OMP (oh-my-pi) with opencode-go subscription using deepseek-v4-flash-0731 (and mimo-v2.5 as advisor). Used pi before (still good, but too minimalist for all general use; works on Android under Termux btw) and OpenCode (which is ok too). I've used Codex the longest, but the $20 sub doesn't last that long. Before that Claude (and Claude on the web - about Nov'25), but nowadays with the new models not possible to do", "context": "", "use_case": "local inference on Apple silicon", "switched_from": "", "story": "llama.cpp"}, {"id": 788243522, "item_id": 49265220, "url": "https://news.ycombinator.com/item?id=49265220", "by": "mark_l_watson", "date": "2026-08-11", "stance": "recommend", "quote": "runs so well on Apple Silicon using MLX!", "excerpt": "I love the wave of new small model releases. Pleasantly surprising that an NVIDIA model runs so well on Apple Silicon using MLX! I was using nemotron-3.5-lightning:30b-mlx with OpenCode on my old (cheap) Mac this morning and no bad experiences except for running slowly.", "context": "", "use_case": "running models on Apple Silicon", "switched_from": "", "story": "Nvidia Nemotron 3.5 lightning and Nemo Switchyard"}, {"id": 776141218, "item_id": 48508826, "url": "https://news.ycombinator.com/item?id=48508826", "by": "LoganDark", "date": "2026-06-12", "stance": "compare", "quote": "still faster than llama.cpp's 70.9t/s and MLX's 80.6t/s with the same model", "excerpt": "I poured a couple days into custom Burn inference for Qwen3-Coder-Next only to find it doesn't come with a speculative decoder, so on my M4 Max I can't push it much further than 120t/s. That's still kinda slow, though still faster than llama.cpp's 70.9t/s and MLX's 80.6t/s with the same model. Claude Fable 5 is recommending I use the Qwen3 MTP -- I worry that will compromise the quality somewhat, but might give it a ", "context": "", "use_case": "local model inference", "switched_from": "", "story": "How to Setup a Local Coding Agent on macOS"}, {"id": 775316177, "item_id": 48457261, "url": "https://news.ycombinator.com/item?id=48457261", "by": "wahnfrieden", "date": "2026-06-08", "stance": "warn", "quote": "MLX is not for end user deployment.", "excerpt": "I use CoreML for models designed for other platforms. I port the models to it but it works for that without much trouble. MLX is not for end user deployment.", "context": "", "use_case": "", "switched_from": "", "story": "Apple Core AI Framework"}, {"id": 769439044, "item_id": 48089940, "url": "https://news.ycombinator.com/item?id=48089940", "by": "quacker", "date": "2026-05-10", "stance": "warn", "quote": "I tried 4bit MLX quants of Qwen 3.5 9B but it eventually would crash due to insufficient memory.", "excerpt": "I could have used this article before I spent the weekend arriving to the same conclusion! Same laptop, and my contrived test was having it fix 50 or so lint errors in a small vibe-coded C++ repo. I wanted it to be able to handle a bunch of small tasks without getting stuck too often. GPT OSS 20B was usable but slow, and actually frequently made mistakes like adding or duplicating statements unnecessarily, listing th", "context": "", "use_case": "running local models", "switched_from": "", "story": "Running local models on an M4 with 24GB memory"}, {"id": 764645009, "item_id": 47790313, "url": "https://news.ycombinator.com/item?id=47790313", "by": "song", "date": "2026-04-16", "stance": "neutral", "quote": "supports mlx for acceleration", "excerpt": "So, on a mac, what good alternative to ollama supports mlx for acceleration? My main use case is that I have an old m1 max macbook pro with 64 gb ram that I use as a model server.", "context": "", "use_case": "Apple Silicon acceleration", "switched_from": "", "story": "Stop Using Ollama"}, {"id": 756700337, "item_id": 47293771, "url": "https://news.ycombinator.com/item?id=47293771", "by": "LoganDark", "date": "2026-03-07", "stance": "neutral", "quote": "port a diffusion language model to Swift+MLX so that I could run it on macOS", "excerpt": "I love Claude Code because it lets me try out ideas a lot faster. Sometimes I have to fight with it to get the outcome I want, but other times it's like I can do something in 3 days that'd otherwise take me weeks. For example, I recently had it port a diffusion language model to Swift+MLX so that I could run it on macOS. It messed up one of the details at first, but it was able to figure out the problem and fix it. A", "context": "", "use_case": "running models on macOS", "switched_from": "", "story": "I'm 60 years old. Claude Code has ignited a passion again"}, {"id": 756437633, "item_id": 47277352, "url": "https://news.ycombinator.com/item?id=47277352", "by": "jmanhype", "date": "2026-03-06", "stance": "neutral", "quote": "integration with MLX for hybrid GPU+ANE training", "excerpt": "I'm not an ML engineer. I used Claude Code (Opus 4.6) to get LoRA fine-tuning gradients running on Apple's Neural Engine \u2014 the dedicated ML chip in every Apple Silicon Mac that has no public training API. 192 gradient dispatches, zero GPU fallbacks, converging loss, all at ~2.8W. Three discoveries found through iteration on real hardware: (1) ANE's matmul op compiles but never executes \u2014 everything must be rewritten ", "context": "", "use_case": "hybrid GPU+ANE training", "switched_from": "", "story": "Show HN: LoRA gradients on Apple's Neural Engine at 2.8W"}, {"id": 749692864, "item_id": 46855804, "url": "https://news.ycombinator.com/item?id=46855804", "by": "llm_nerd", "date": "2026-02-02", "stance": "neutral", "quote": "MLX tries to really leverage platform optimizations", "excerpt": "Sure, I directly and explicitly talked about Apple's version of tensor cores in the GPU. But the ANE is by every definition a neural accelerator. Yes, I'm aware of Apple's weird branding for their tensor cores. \"In fact MLX does not even support ANE yet\" I didn't say otherwise. The ANE is a fantastic unit for small, power-efficient models, like extracting text from images, doing depth modelling, etc. It's not made fo", "context": "", "use_case": "", "switched_from": "", "story": "My thousand dollar iPhone can't do math"}, {"id": 746082881, "item_id": 46630180, "url": "https://news.ycombinator.com/item?id=46630180", "by": "__jf__", "date": "2026-01-15", "stance": "neutral", "quote": "only had to replace Transformers with Apple's MLX python library", "excerpt": "For vector generation I started using Meta-LLama-3-8B in april 2024 with Python and Transformers for each text chunk on an RTX-A6000. Wow that thing was fast but noisy and also burns 500W. So a year ago I switched to an M1 Ultra and only had to replace Transformers with Apple's MLX python library. Approximately the same speed but less heat and noise. The Llama model has 4k dimensions so at fp16 thats 8 kilobyte per c", "context": "runs local RAG", "use_case": "running models on Apple silicon", "switched_from": "Transformers", "story": "Ask HN: How are you doing RAG locally?"}]}