{"slug": "kimi-k3-how-an-open-weights-model-reached-the-level-of-claude-and-gpts-best", "title": "Kimi K3: How an Open-Weights Model Reached the Level of Claude and GPT’s Best Models", "summary": "Moonshot AI released Kimi K3, an open-weights model with 2.78 trillion total parameters and 104.2 billion active per token, achieving a composite Intelligence Index score of 57 on Artificial Analysis, placing it in serious contention with frontier models like Claude and GPT. The model's improvements stem from architectural changes, including a new long-context memory design and expanded expert scaling, rather than a single breakthrough. Moonshot's technical report highlights a significant performance gain on an internal camera-repair task, though the company notes the result is task-specific and not generalizable.", "body_md": "A model whose weights you can download is now part of the performance conversation that usually belongs to Claude and GPT.\n\nThat is what makes Kimi K3 exciting.\n\nIt does **not** mean K3 wins every benchmark or matches every\nClaude and GPT configuration. The title means something narrower and still\nimportant: this open-weights release is close enough to be compared seriously\nwith current frontier systems, even though it does not lead every test. When I\nchecked\n[Artificial Analysis’ K3 (max) page](https://artificialanalysis.ai/models/kimi-k3)\non August 3, 2026, its composite Intelligence Index score was 57. That is not\nthe top score on every task. It is a serious top-tier signal from a model\nwhose weights are public.\n\nSo how did K3 make such a large leap from Kimi K2.6?\n\nThe short answer is: **not through one magic trick.** Moonshot\nchanged the usable size of the model, how information moves through it, the\nlong tasks it trained on, the environments in which it practiced tool use, the\namount of reasoning it can spend, and the systems needed to run all of that.\n\nThat matters because the frontier is increasingly a system story, not only a\nmodel story. In my earlier post,\n[GPT-5.5 Is Not Just a Better Model. It Is a Deployment Story](https://eido-askayo.blogspot.com/2026/04/gpt-5-dot-5-is-not-just-better-model.html), I argued that the surrounding harness matters once models take longer,\ntool-driven actions. K3 makes the same point from the open-weights side.\n\n“Open weights” means the learned model files can be downloaded and run or adapted outside Moonshot’s hosted product. That is a meaningful difference from a closed API-only model. It gives researchers and companies more control over where the model runs and how it is integrated.\n\nBut it does not mean the entire recipe is public. The training data, full\ntraining stack, internal environments, and production fleet are not handed\nover with the checkpoint. And K3 is licensed under the\n[Kimi K3 License](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE), not a simple unrestricted MIT license. Larger commercial deployments have\nextra conditions.\n\nThere is also a practical constraint: K3 is huge. Moonshot lists 2.8 trillion total parameters, with 104 billion active for each token. Downloadable does not mean “easy to run on a laptop.”\n\nStill, an open-weights model at this scale changes who can inspect, host, test, and build around it. That is the real hook.\n\nMoonshot’s\n[K3 technical report](https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf)\nincludes a chart that is hard to ignore. In one internal camera-repair task,\nK3 ends at `1.000`\n\nverifier-assessed completion. K2.6 ends at\n`0.560`\n\n.\n\nThat looks enormous. It is also only one task.\n\nThe chart does not show a public task set, a run count, or an experiment that turns off one K3 feature at a time. It cannot tell us that K3 is “twice as good” in general. What it can tell us is more specific: Moonshot’s own evidence points to a large difference in one long-horizon, tool-driven task.\n\nThat distinction matters. A model result is always a result from a particular model, task, tools, context limit, reasoning budget, and scoring method. Change the setup and the number can change too.\n\nK2.6 was already a serious sparse model. It had 1 trillion total parameters, 32 billion active parameters, native multimodality, a 256K-token context window, and agentic coding demonstrations.\n\nK3 expands that base, but the deeper change is not only scale.\n\n| What changed | K2.6 public disclosure | K3 disclosure | What it may help with |\n|---|---|---|---|\n| Total model size | 1T parameters | 2.78T parameters | More stored capacity |\n| Active parameters per token | 32B | 104.2B | More capacity used for each decision |\n| Expert scale | 384 routed experts, 8 active, 1 shared expert | 896 routed experts, 16 active, 2 shared experts | A wider pool of specialist computation |\n| Sequence memory | MLA attention | New long-context memory design (hybrid KDA plus gated MLA) | A compact running state plus periodic global interaction |\n| Reusing earlier processing | No Attention Residuals documented | Block Attention Residuals | Later layers can selectively reuse earlier representations |\n| Expert design and routing | SwiGLU; no extra routing method disclosed | New specialist-team design and balancing (Stable LatentMoE, SiTU-GLU, Quantile Balancing) | Less expert traffic and more stable, balanced routing |\n| Long-context training | 256K-token context | 1M context plus staged long-context training | Longer work with evidence spread across a session |\n| Vision recipe | 400M MoonViT | From-scratch MoonViT-V2, about 401M parameters | A redesigned multimodal training path |\n\nThe “not disclosed” cells do not mean K2.6 definitely lacked the feature. They mean Moonshot’s public K2.6 materials do not name it. That is the strongest honest comparison we can make.\n\nThink of an MoE as a company with many specialist teams. Its router wakes only the relevant teams for each token. K3 has more teams and uses more of them per decision. And for a long project, its hybrid attention works like a running notebook plus periodic full-room meetings. KDA provides the compact state; gated MLA can reconnect distant parts of the context. Attention Residuals and Stable LatentMoE give K3 documented mechanisms to selectively reuse earlier representations and reduce specialist communication.\n\nThe simple mental model is this: K3 is not just a bigger brain. Moonshot redesigned the roads between its memory, layers, and specialist workers.\n\nMoonshot’s scaling-law figure reports about\n**2.5x higher scaling efficiency** for the full K3 recipe than\nK2. Notice the careful wording: full recipe, and K2 rather than K2.6. The\nfigure moves many levers at once. It is evidence that the combined design\nscaled better. It is not proof that one architecture feature caused the later\nK2.6-to-K3 gap.\n\nThe biggest misconception would be to say K2.6 did not do agentic work. It did. Moonshot’s K2.6 material already describes long-running coding tasks, tool calls, context management, and Agent Swarm behavior.\n\nK3 pushes further in three directions.\n\nFirst, it was trained for a longer working memory. The report describes a progression from shorter contexts to 1M tokens, plus long documents, videos, and synthetic tasks where the needed evidence is spread across the full context. A larger context limit only matters if the model learns how to use it. Moonshot’s claim is that K3’s architecture, data, and curriculum were designed together for that purpose.\n\nSecond, it practiced in more realistic environments. The report describes coding, search, professional-workflow, visual, kernel-optimization, and persistent mock-app tasks. Some use verifiers that can check whether an answer actually works. Some require many tool calls and large contexts. Think of this as training a worker to finish full projects, not only a student to answer isolated exam questions.\n\nThird, Moonshot trained different reasoning budgets. It created nine specialist teachers: general, general-agent, and coding-agent models, each at low, high, and max effort. It then consolidated them into one K3 model using Multi-Teacher On-Policy Distillation.\n\nIn plain English: instead of training one behavior for every task, Moonshot trained specialists for different kinds of work and different amounts of thinking, then taught one model to bring those behaviors together.\n\nThis is a plausible explanation for K3’s stronger long-horizon behavior. It is not an isolated proof. More scale, different data, better environments, and more inference-time reasoning all changed at once.\n\nSome K3 work is easy to miss because it does not look like a benchmark headline. Think of it as the workshop around a much bigger machine.\n\nMoonshot tuned K3 to work with lower-precision expert weights, trained a fast helper model to propose likely next tokens, reused memory more efficiently, and built many sandboxed environments for training and testing. The technical names are quantization-aware post-training, speculative decoding, cache design, and agent sandboxes.\n\nThese are not direct explanations for a higher reasoning score. They are\n**enablers**. They can make it possible to train, evaluate, and\nserve a much larger model on longer tool-based tasks without making the\nprocess impossible to afford or operate.\n\nThat is why “the model” is an incomplete explanation here. A 1M-token, 2.8T-parameter agent needs the surrounding machinery to be useful in practice.\n\n```\nflowchart TB\n    recipe[\"K3's disclosed recipe<br/>capacity, redesign, long-context training,<br/>tool environments, reasoning budgets\"] -. \"reported with\" .-> results[\"K3's reported task results\"]\n    setup[\"Test setup<br/>tasks, tools, limits, model version\"] -->|\"can change\"| results\n    serving[\"Serving and cost work\"] --> practical[\"Practical deployment\"]\n```\n\n*Moonshot disclosed all of these changes. It did not publish an experiment\nshowing how much each change caused the jump.*\n\nThe strong version of this story is not “K3 crushed everything.” The strong version is that Moonshot disclosed a surprisingly broad program behind an open-weights frontier release.\n\nIts published results still need careful reading. K3’s main report uses max reasoning effort. Different coding tasks use Kimi Code, Claude Code, Codex, or other harnesses. Some competitor values come from public leaderboards, some from other companies, and some from Moonshot’s own runs. Tool access, context management, fallbacks, refusals, and task versions can change an outcome.\n\nArtificial Analysis adds useful independent context, but its 57 score is also a composite across nine evaluations. It is not a universal answer to “which model is best?”\n\nThere is another gap. The K2.6 and K3 source sets I reviewed do not include a dedicated standalone system card. K3’s technical report does contain some cyber evaluation and agent-control material, but that is not the same as a complete deployment-safety account. Anyone using an agent with real tools still needs least-privilege access, sandboxing, logs, confirmation gates, and evaluation in their own environment.\n\nK3 did not jump ahead because Moonshot found a single secret switch.\n\nIt built a substantially larger active model, changed its architecture, trained it over longer contexts and more tool-rich environments, taught it to use different reasoning budgets, and built the systems that make that work practical.\n\nThat is why this release is worth watching. The exciting part is not only that K3 has strong numbers. It is that an open-weights model can now arrive with a recipe broad enough to challenge the old assumption that frontier performance must stay inside a closed API.", "url": "https://wpnews.pro/news/kimi-k3-how-an-open-weights-model-reached-the-level-of-claude-and-gpts-best", "canonical_source": "https://eido-askayo.blogspot.com/2026/08/kimi-k3-how-open-weights-model-reached.html", "published_at": "2026-08-05 07:26:07+00:00", "updated_at": "2026-08-05 07:35:33.546016+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "ai-research", "ai-products"], "entities": ["Moonshot AI", "Kimi K3", "Kimi K2.6", "Artificial Analysis", "Claude", "GPT"], "alternates": {"html": "https://wpnews.pro/news/kimi-k3-how-an-open-weights-model-reached-the-level-of-claude-and-gpts-best", "markdown": "https://wpnews.pro/news/kimi-k3-how-an-open-weights-model-reached-the-level-of-claude-and-gpts-best.md", "text": "https://wpnews.pro/news/kimi-k3-how-an-open-weights-model-reached-the-level-of-claude-and-gpts-best.txt", "jsonld": "https://wpnews.pro/news/kimi-k3-how-an-open-weights-model-reached-the-level-of-claude-and-gpts-best.jsonld"}}