Moonshot AI published the weights for Kimi K3 on Hugging Face today, July 27, 2026, eleven days after the model went live through its API. At 2.8 trillion total parameters it is the largest open-weight model released to date, and the first in the three-trillion-parameter class.
That much is straightforward. Most of what has been written about this release since is not. Two claims circulating widely are both wrong, and both are checkable against the model card in under a minute: that K3 ships under a Modified MIT license, and that it activates roughly 50 billion parameters per token. This article works from primary sources: Moonshot's model card and technical blog, the independent evaluation from Artificial Analysis, and the engineering notes from vLLM and the providers hosting it.
Key Takeaways
- K3 activates 104B of 2.8T parameters per token (Moonshot model card), not the ~50B figure circulating in community write-ups
- It scores 57 on the Artificial Analysis Intelligence Index, ranking #3, behind Claude Fable 5 and GPT-5.6 Sol (Artificial Analysis)
- Its hallucination rate regressed to 51%, up from K2.6's 39% (Artificial Analysis), the clearest regression in the release
- The license is the Kimi K3 License, not MIT: Model-as-a-Service providers over $20M annual revenue need a separate agreement with Moonshot
- Weights are ~1.4TB, with Moonshot recommending 64+ accelerators: downloadable is not the same as runnable
K3 is a mixture-of-experts model with a native vision encoder and a 1,048,576-token context window. The model card gives the full configuration:
| Property | Value |
|---|---|
| Total parameters | 2.8T |
| Activated parameters per token | 104B |
| Layers | 93 |
| Attention composition | 69 KDA + 24 Gated MLA |
| Experts | 896 (16 selected per token, 2 shared) |
| Vision encoder | MoonViT-V2 (401M) |
| Vocabulary | 160K |
| Context length | 1,048,576 tokens |
| Quantization | MXFP4 weights / MXFP8 activations |
Three architectural choices are worth naming. Kimi Delta Attention (KDA) covers 69 of the 93 layers, with Gated MLA handling the remaining 24, a hybrid rather than a uniform attention stack. Stable LatentMoE handles routing across the 896 experts. And the quantization is not a post-training conversion: Moonshot states the model went through quantization-aware training from the SFT stage onward, with MXFP4 weights and MXFP8 activations trained in rather than fitted afterward.
That last detail is the one with practical consequences. A 2.8T-parameter model at FP16 would be unservable outside a handful of datacenters. Training directly in a 4-bit format is what brings the release down to roughly 1.4TB and makes third-party hosting viable at all.
A widely-shared Hugging Face community post estimated K3's active parameters at "~50B equivalent (16/896 experts per token)". That derivation is wrong. It scales total parameters by the expert ratio and ignores the 2 shared experts plus the dense attention stack that runs on every token. Moonshot's own model card states 104B activated. Use that number.
Benchmark numbers for this release come from two sources that do not fully agree, so they are worth keeping separate.
Artificial Analysis runs its own harness and publishes cost and token-consumption data alongside scores:
| Metric | Kimi K3 | Context |
|---|---|---|
| Intelligence Index | 57 (rank #3) | Comparable to Opus 4.8 and GPT-5.5; behind Fable 5 and GPT-5.6 Sol |
| GDPval-AA v2 (Elo) | 1668 | K2.6 scored 1190. Fable 5 1760, Opus 4.8 1600, GLM-5.2 1514, GPT-5.5 1494 |
| AA-Briefcase (Elo) | 1547 | +732 over K2.6; second only to Fable 5 |
| AutomationBench-AA | 53% | Ranked #1 |
| Cost per task | $0.94 | GPT-5.6 Sol $1.04, Opus 4.8 $1.80, GLM-5.2 $0.32, DeepSeek V4 Pro $0.04 |
| Output tokens to run index | 132M | 21% fewer than K2.6's 166M |
| Hallucination rate | 51% | K2.6 was 39% |
The efficiency result is the underrated one. K3 used 21% fewer output tokens than its predecessor while scoring substantially higher, which is the opposite of the usual pattern where reasoning models buy accuracy with token volume.
Moonshot's model card publishes a comparison across more than 40 benchmarks. Selected results, with K3 against Claude Fable 5, GPT-5.6 Sol and Claude Opus 4.8:
| Benchmark | Kimi K3 | Fable 5 | GPT-5.6 Sol | Opus 4.8 |
|---|---|---|---|---|
| GPQA Diamond | 93.5 | 92.6 | 94.1 | 91.0 |
| Terminal-Bench 2.1 | 88.3 | 88.0 | 88.8 | 84.6 |
| DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 |
| SWE-Marathon | 42.0 | 35.0 | 39.0 | 40.0 |
| Kimi Code Bench 2.0 | 72.9 | 76.9 | 64.8 | 71.7 |
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 |
| MCPMark-Verified | 94.5 | 87.4 | 92.9 | 76.4 |
| HLE-Full | 43.5 | 53.3 | 44.5 | 49.8 |
| OSWorld 2.0 | 58.3 | 66.1 | 62.6 | 55.7 |
| τ³-Banking | 33.4 | 26.8 | 33.0 | 27.6 |
| Harvey Lab-AA | 94.6 | 93.6 | 87.2 | 91.1 |
| OmniDocBench | 91.1 | 89.8 | 85.8 | 87.9 |
K3 genuinely leads on tool-use and search-heavy work: BrowseComp, MCPMark-Verified, DeepSearchQA, AutomationBench, and long-context retrieval via AA-LCR. The τ³-Banking and Harvey Lab-AA results suggest it holds up on domain-constrained agentic tasks too.
Moonshot reports a GDPval-AA v2 Elo of 1686. Artificial Analysis measured 1668 on the same benchmark. On AA-Briefcase the gap is a single point: 1548 vendor-reported against 1547 independently measured. Moonshot's own blog post also cites DeepSWE at 67.3 where its model card says 67.5.
None of these gaps is large enough to change the ranking, and the near-identical AA-Briefcase figures suggest the evaluation methodology is broadly reproducible. But they are a reminder that vendor tables and independent harnesses are not the same measurement, and a release this heavily benchmarked deserves the distinction being made explicitly.
💡 Created originally by OpenAI, GDPval measures how well AI models perform on real-world, economically valuable knowledge-work tasks.
Instead of short academic Q&A, it tests complex tasks across 44 occupations in 9 major industries (finance, legal, consulting, engineering, healthcare, etc.).
These tasks reflect actual deliverables produced by industry professionals (e.g., writing research reports, financial modeling, workflow planning).
The Score: Elo of 1686:
Elo System: Rather than a simple percentage accuracy, models earn an Elo rating based on head-to-head pairwise comparisons of their work deliverables.
Human Baseline: The Elo scale is anchored to a standard human expert baseline of 1,000 Elo.
Interpretation: A score of 1686 Elo places the model significantly above the human baseline, indicating top-tier frontier performance on long-horizon professional work.
The hallucination regression is the most significant negative finding, and Moonshot's own materials do not address it. Artificial Analysis measured a 51% hallucination rate against K2.6's 39%. A model that is better at agentic execution while being measurably less reliable about facts is a specific and awkward tradeoff. It is exactly the profile you do not want for research or summarization work, and exactly where the marketing pressure will push it.
The vendor's own comparison table also shows K3 losing to Claude Fable 5 on a substantial share of benchmarks, including FrontierSWE (81.2 vs 86.6), DeepSWE, HLE-Full (43.5 vs 53.3), OSWorld 2.0, PostTrainBench, OfficeQA Pro and Legal Research Bench. Most notably it loses on Kimi Code Bench 2.0, Moonshot's own benchmark, scoring 72.9 against Fable 5's 76.9. Publishing that is to Moonshot's credit. It also undercuts the framing that K3 has reached parity with the frontier.
Then there is price. K3 costs $3.00 per million input tokens and $15.00 per million output tokens, with cache hits at $0.30. Kimi K2.6 was $0.95 and $4.00. That is roughly 3x on input and 3.75x on output for the same vendor's flagship, one generation apart.
Compounding that, K3 launched with a single reasoning effort level. Moonshot's blog states the model runs at "max thinking effort by default, with low- and high-effort modes to be introduced in subsequent updates". Simon Willison found this made trivial prompts expensive (a single SVG generation test cost 25 cents), and observed that a short prompt registered 95 tokens, implying an undocumented system prompt of around 85 tokens on every call.
K3 ships under the Kimi K3 License. Hugging Face categorizes it as other, not as an SPDX-recognized MIT identifier. Reports that this is a "Modified MIT license" are incorrect. That was the license on K2.7 Code, a different model.
The grant itself is broad and MIT-like: use, copy, modify, distribute, sublicense, sell, and run, deploy, fine-tune or build derivative works from the weights. Two commercial gates sit on top (license analysis via Unite.AI):
Neither applies to internal use, or to access through Moonshot's own products and certified inference partners. Researcher Nathan Lambert characterized the terms publicly as "inspired by MIT but distinctly non-commercial".
For most readers (internal deployment, research, fine-tuning, products below those thresholds), this is functionally permissive. For anyone building a hosting business on it, it is not open source in the OSI sense, and the distinction matters before legal review, not after. Do not tell your legal team this is MIT-licensed.
Probably not, and this is where the open-weights story gets thin.
The weights are approximately 1.4TB even at MXFP4. Moonshot recommends deploying on supernode configurations of 64 or more accelerators. The vLLM team validated a 16-GPU DP16+EP16 configuration, which is the smallest credible setup published so far. Sixteen datacenter GPUs is not a homelab.
Supported inference engines are vLLM, SGLang and TokenSpeed. vLLM shipped day-0 support with dedicated NVIDIA kernels for KDA decode and attention residuals, plus AMD ROCm support via a FlyDSL MoE kernel. Native 1M context, vision, tool calling, structured output and disaggregated prefill/decode all work from day one. MXFP4 runs natively on NVIDIA Blackwell and AMD MI400.
So the release is genuinely useful to organizations with GPU clusters, sovereignty requirements, or a need to fine-tune. For an individual developer, "open weights" here means you can inspect and audit the model, and you can choose between hosts. It does not mean you can run it.
Confirmed as of July 27, 2026:
| Provider | Status |
|---|---|
| Moonshot first-party | Live since July 16: kimi.com, Kimi Work, Kimi Code, Kimi API |
| OpenRouter | Live, routing across third-party providers, $3/$15 per 1M tokens |
| Fireworks AI | Day-0 inference and training; serverless standard and priority tiers, batch mode at 50% discount, US-only zero-data-retention endpoints |
| Together AI | Day-0 hosting |
I could not confirm live K3 endpoints on Baseten, Groq, DeepInfra or Novita at the time of writing. Several roundups list them; none that I checked cite a working endpoint. Treat those as unverified rather than available.
K3 is a strong fit for agentic and tool-use workloads (browsing, MCP-based tool calling, long-horizon automation, document understanding) where its benchmark leads are real and independently corroborated by AutomationBench and AA-Briefcase. At $0.94 per task on the Artificial Analysis index it is roughly half the cost of Opus 4.8 for comparable index-level intelligence.
It is a poor fit for anything where factual reliability is the primary requirement, given a 51% hallucination rate that moved in the wrong direction. It is also poorly suited to high-volume simple queries until the promised low-effort reasoning mode ships.
The more durable point is about the gap. An open-weight model now sits at #3 on an independent intelligence index, above one frontier lab's previous flagship. That gap is measured in months, and the weights are on Hugging Face today.
Is Kimi K3 open source? No. K3 ships under the Kimi K3 License, which Hugging Face lists as other. The grant is MIT-like, but Model-as-a-Service providers exceeding $20M in annual revenue need a separate agreement with Moonshot. It is open weights, not OSI-approved open source.
How many parameters does Kimi K3 activate? 104 billion of 2.8 trillion total, per Moonshot's model card. The routing selects 16 of 896 experts per token, plus 2 shared experts. Community estimates of ~50B are incorrect. They omit shared experts and the dense attention stack.
How does Kimi K3 compare to Claude Fable 5? Artificial Analysis ranks K3 #3 on its Intelligence Index at 57, behind Fable 5. Moonshot's own table shows K3 losing to Fable 5 on FrontierSWE, DeepSWE, HLE-Full and its own Kimi Code Bench 2.0, while leading on BrowseComp, MCPMark and SWE-Marathon.
What hardware do you need to run Kimi K3? Moonshot recommends 64 or more accelerators. vLLM validated a 16-GPU DP16+EP16 configuration as a smaller option. The weights are roughly 1.4TB at MXFP4 quantization. Self-hosting is not realistic below datacenter-scale GPU capacity.
Where can I use Kimi K3 today? Moonshot's own API and apps, OpenRouter, Fireworks AI, and Together AI all serve K3 as of July 27, 2026. Moonshot's API pricing is $3.00 per million input tokens and $15.00 per million output, with cache hits at $0.30.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。