--- title: Kimi K3 matches GPT-5 on benchmarks, costs 4x less, and... description: Moonshot AI has released the full weights for Kimi K3 on Hugging Face — a 2.8-trillion-parameter Mixture-of-Experts model that activates 104B parameters per... canonical: https://daily.dev/posts/kimi-k3-matches-gpt-5-on-benchmarks-costs-4x-less-and-the-weights-are-free-g7buv7gcl twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Kimi K3 matches GPT-5 on benchmarks, costs 4x less, and the weights are free | daily.dev og:description: Moonshot AI has released the full weights for Kimi K3 on Hugging Face — a 2.8-trillion-parameter Mixture-of-Experts model that activates 104B parameters per... og:url: https://daily.dev/posts/kimi-k3-matches-gpt-5-on-benchmarks-costs-4x-less-and-the-weights-are-free-g7buv7gcl og:image: https://api.daily.dev/og/posts/g7bUV7GCl.png og:image:alt: Kimi K3 matches GPT-5 on benchmarks, costs 4x less, and the weights are free og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Kimi K3 matches GPT-5 on benchmarks, costs 4x less, and the weights are free **[Trends](https://daily.dev/sources/trends)** · 2 min read · 50 upvotes · 10 comments ## Summary Moonshot AI has released the full weights for Kimi K3 on Hugging Face — a 2.8-trillion-parameter Mixture-of-Experts model that activates 104B parameters per request, compressed to ~1.56TB in MXFP4. It benchmarks alongside Claude and GPT-5 while costing roughly 4x less via API. GitHub Copilot has already integrated it across all plan tiers (Pro through Enterprise), hosted on Fireworks AI. Hands-on testing across 10 Blender and Godot tasks showed K3 performing on par with or better than Claude and GPT-4o. Self-hosting is technically possible but impractical — one community effort achieved only 0.32–0.34 tokens/sec on a 64GB MacBook Pro due to SSD streaming bottlenecks. The broader takeaway is that the quality gap between open-weight and closed frontier models continues to narrow, making the cost and compliance trade-offs increasingly favorable for open models. ## Content Moonshot AI dropped the full 2.8-trillion-parameter Kimi K3 on Hugging Face, and the community is split between genuine excitement and a reality check about what "open weights" actually means in practice. The model itself is legitimately impressive. It's a Mixture-of-Experts architecture that only activates 104B parameters per request, compressed to ~1.56TB in MXFP4. Benchmarks put it competitive with Claude and GPT-5 on coding and reasoning, at a fraction of the API cost. One hands-on comparison across 10 Blender and game dev tasks found it performing on par with or better than Claude and GPT-4o while being nearly 4x cheaper. The architecture is doing some genuinely interesting things to stay stable at this scale: a latent compression step before routing to 896 fine-grained experts (16 active per token), RMS normalization to prevent activation overflow, a bounded activation function called SiTU-GLU, and a quantile balancing algorithm that skips auxiliary losses entirely for load balancing. That last one is the kind of engineering detail that makes ML people stop scrolling. But here's where the discourse gets honest. Someone did run it on a 64GB MacBook Pro. Throughput: 0.32–0.34 tokens per second, bottlenecked by SSD streaming. A separate community project called `kimi-k3-in-c` — a 176KB pure-C99 engine — claims to run the model with only 8.24GB of RAM and no GPU. The reaction to that has been equal parts impressed and skeptical about what's actually being quantized away to get there. At the other end of the hardware spectrum, someone ran the full model on a 16x GB10 cluster and hit 750 tokens per second peak throughput. That's the gap. Open weights, closed power. Meanwhile, GitHub Copilot quietly added Kimi K3 across all plan tiers (Pro through Enterprise), hosted on Fireworks AI. Business and Enterprise admins have to explicitly enable it, with a note to review open-weight model policies for compliance. That last line is doing a lot of work — it's a reminder that "open" and "enterprise-ready" are still two different conversations. The leaderboard position is also in motion. K3 currently leads open-weight models at 57 on the relevant benchmark, one point ahead of Qwen3.8 Max — which Alibaba hasn't fully released yet. That lead may not last the week. ## Questions this post answers ### How many parameters does Kimi K3 activate per request compared to its total size? Kimi K3 has 2.8 trillion total parameters but activates only 104 billion per request, using a Mixture-of-Experts architecture with 896 fine-grained experts and 16 active per token. The full model is compressed to roughly 1.56TB in MXFP4 format. This sparse activation is what lets it compete with much larger dense models on coding and reasoning benchmarks. _Teams weighing MoE efficiency against dense models can track model architecture comparisons like this on daily.dev._ ### Can you run Kimi K3 locally on a laptop without a GPU? Yes, with major throughput tradeoffs. On a 64GB MacBook Pro, Kimi K3 ran at only 0.32-0.34 tokens per second, bottlenecked by SSD streaming. A community project called kimi-k3-in-c, a 176KB pure-C99 engine, claims to run it using just 8.24GB of RAM with no GPU, though the community remains split on what quantization tradeoffs make that possible. _Developers testing local inference limits for huge open models can follow real-world benchmarks like this on daily.dev._ ### Is Kimi K3 available in GitHub Copilot? Yes, GitHub Copilot added Kimi K3 across all plan tiers from Pro through Enterprise, hosted via Fireworks AI. Business and Enterprise administrators must explicitly enable it and are advised to review their open-weight model policies for compliance before doing so, since enterprise governance around open-weight models differs from standard vendor-hosted models. _Engineering leads choosing AI coding models under enterprise policy can track rollouts like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 0 discussions (as of 2026-09-13). **TL;DR:** No external discussion threads were provided, so no community signal is available to analyze for this post. **Sentiment:** 50% positive · 30% mixed · 20% skeptical ## Community discussion Top comments from developers on daily.dev. **@aaronschiers** · 4 upvotes > i cannot be the only one that thinks AI Generated blog posts like this are besides the point of this platform as a whole right ? i mean whats the point of having people post things if you can just offload it to an AI. report on the topic all you want, but at least use someones actual creativity to write the posts..... **@baz14** · 2 upvotes > Open weights are a mandatory requirement for further development of LLM models. Also, if model can be installed locally, there is no dependency on a single provider, reducing the risk of tool loss. > > The next step is opening up the training data. Without this, it is impossible to verify whether it was obtained legally or licensed, preventing the results from being used as legitimate. **@fahimahammed** · 1 upvotes > Impressive, but benchmarks don’t tell the whole story. Free weights don’t mean cheap inference. Still, the progress of open models is exciting. 🚀 **@hemna** · 1 upvotes > sure, but can't run it at home. **@bob\_marley\_556** · 0 upvotes > how to use this tool in fedora system --- Tags: [#open-source](https://daily.dev/tags/open-source), [#llm](https://daily.dev/tags/llm), [#github](https://daily.dev/tags/github), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts), [#kimi](https://daily.dev/tags/kimi) [View this post on daily.dev](https://daily.dev/posts/kimi-k3-matches-gpt-5-on-benchmarks-costs-4x-less-and-the-weights-are-free-g7buv7gcl) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Kimi K3 matches GPT-5 on benchmarks, costs 4x less, and the weights are free","url":"https://daily.dev/posts/kimi-k3-matches-gpt-5-on-benchmarks-costs-4x-less-and-the-weights-are-free-g7buv7gcl","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/kimi-k3-matches-gpt-5-on-benchmarks-costs-4x-less-and-the-weights-are-free-g7buv7gcl"},"datePublished":"2026-08-06T17:58:48.973Z","dateModified":"2026-09-13T20:00:19.359Z","description":"Moonshot AI has released the full weights for Kimi K3 on Hugging Face — a 2.8-trillion-parameter Mixture-of-Experts model that activates 104B parameters per...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/5fccc92b5d9fc088b260ff357b9dc569?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/5fccc92b5d9fc088b260ff357b9dc569?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":10,"discussionUrl":"https://daily.dev/posts/kimi-k3-matches-gpt-5-on-benchmarks-costs-4x-less-and-the-weights-are-free-g7buv7gcl","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":50},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":10}],"keywords":"open-source,llm,github,mixture-of-experts,kimi","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Kimi K3 matches GPT-5 on benchmarks, costs 4x less, and the weights are free"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/kimi-k3-matches-gpt-5-on-benchmarks-costs-4x-less-and-the-weights-are-free-g7buv7gcl","comment":[{"@type":"Comment","text":"i cannot be the only one that thinks AI Generated blog posts like this are besides the point of this platform as a whole right ? i mean whats the point of having people post things if you can just offload it to an AI. report on the topic all you want, but at least use someones actual creativity to write the posts…","datePublished":"2026-08-07T08:40:49.635Z","url":"https://daily.dev/posts/g7bUV7GCl#c-43N3GSVIH","author":{"@type":"Person","name":"Aaron Schiers","url":"https://daily.dev/aaronschiers","image":"https://avatars.githubusercontent.com/u/142320692?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4}},{"@type":"Comment","text":"Open weights are a mandatory requirement for further development of LLM models. Also, if model can be installed locally, there is no dependency on a single provider, reducing the risk of tool loss.\nThe next step is opening up the training data. Without this, it is impossible to verify whether it was obtained legally or licensed, preventing the results from being used as legitimate.","datePublished":"2026-08-07T05:31:43.039Z","url":"https://daily.dev/posts/g7bUV7GCl#c-15fXIjH5g","author":{"@type":"Person","name":"Baz","url":"https://daily.dev/baz14","image":"https://lh3.googleusercontent.com/a/ACg8ocKd2iFU8YS4sV-rBJptuYSF8EHeQnm23N_dcXE1uON7_GMIlZg=s96-c"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2}},{"@type":"Comment","text":"Impressive, but benchmarks don’t tell the whole story. Free weights don’t mean cheap inference. Still, the progress of open models is exciting. 🚀","datePublished":"2026-08-09T04:09:11.950Z","url":"https://daily.dev/posts/g7bUV7GCl#c-eblg0deAb","author":{"@type":"Person","name":"Fahim Ahammed Firoz","url":"https://daily.dev/fahimahammed","image":"https://lh3.googleusercontent.com/a/ACg8ocLPDnjeV4QsaVwi1lXMBgMAKbog9I4fgMTTJnHRKE-Uc1NmSJY0=s96-c"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}},{"@type":"Comment","text":"sure, but can’t run it at home.","datePublished":"2026-08-10T13:38:54.293Z","url":"https://daily.dev/posts/g7bUV7GCl#c-M3hjbslRc","author":{"@type":"Person","name":"Chewbacca","url":"https://daily.dev/hemna","image":"https://avatars.githubusercontent.com/u/672207?v=4"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}},{"@type":"Comment","text":"how to use this tool in fedora system","datePublished":"2026-08-07T04:55:16.025Z","url":"https://daily.dev/posts/g7bUV7GCl#c-cMmQRi6ep","author":{"@type":"Person","name":"Aashish pokhrel","url":"https://daily.dev/bob_marley_556","image":"https://avatars.githubusercontent.com/u/126847212?v=4"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/kimi-k3-matches-gpt-5-on-benchmarks-costs-4x-less-and-the-weights-are-free-g7buv7gcl#faq","mainEntity":[{"@type":"Question","name":"How many parameters does Kimi K3 activate per request compared to its total size?","acceptedAnswer":{"@type":"Answer","text":"Kimi K3 has 2.8 trillion total parameters but activates only 104 billion per request, using a Mixture-of-Experts architecture with 896 fine-grained experts and 16 active per token. The full model is compressed to roughly 1.56TB in MXFP4 format. This sparse activation is what lets it compete with much larger dense models on coding and reasoning benchmarks. Teams weighing MoE efficiency against dense models can track model architecture comparisons like this on daily.dev."}},{"@type":"Question","name":"Can you run Kimi K3 locally on a laptop without a GPU?","acceptedAnswer":{"@type":"Answer","text":"Yes, with major throughput tradeoffs. On a 64GB MacBook Pro, Kimi K3 ran at only 0.32-0.34 tokens per second, bottlenecked by SSD streaming. A community project called kimi-k3-in-c, a 176KB pure-C99 engine, claims to run it using just 8.24GB of RAM with no GPU, though the community remains split on what quantization tradeoffs make that possible. Developers testing local inference limits for huge open models can follow real-world benchmarks like this on daily.dev."}},{"@type":"Question","name":"Is Kimi K3 available in GitHub Copilot?","acceptedAnswer":{"@type":"Answer","text":"Yes, GitHub Copilot added Kimi K3 across all plan tiers from Pro through Enterprise, hosted via Fireworks AI. Business and Enterprise administrators must explicitly enable it and are advised to review their open-weight model policies for compliance before doing so, since enterprise governance around open-weight models differs from standard vendor-hosted models. Engineering leads choosing AI coding models under enterprise policy can track rollouts like this on daily.dev."}}]} ```