On 27 July 2026 at 00:00 UTC, Moonshot AI publishes the full Kimi K3 open weights on Hugging Face under a Modified MIT licence, making a 2.8-trillion-parameter Sparse Mixture-of-Experts model the largest downloadable LLM to date. Independent testing places K3 fourth on the Artificial Analysis Intelligence Index and first on Frontend Code Arena, but the same third-party evaluation clocked its hallucination rate at roughly 51% on AA-Omniscience, a figure Moonshot did not include in its published benchmark charts.
That gap between what the weights can do and what the lab chose to disclose is the story worth reading before you download 594GB of safetensors.
TL;DR
- Kimi K3 open weights ship 27 July 2026, 00:00 UTC, at
huggingface.co/moonshotai, ~594GB in MXFP4 safetensors. - Architecture: 2.8T total parameters, ~50B active per token, 896 experts (16 routed per token), 1,048,576-token context.
- Independent scores: Intelligence Index 57.1 (rank 4), Frontend Code Arena 1,679 pts (rank 1, 76% win rate).
- AA-Omniscience hallucination rate ~51%, up from 39% on K2.6, and not shown in Moonshot's charts.
- Minimum realistic self-host: 8× H100 80GB (experimental), 64+ accelerators recommended; MXFP4 natively supported only on NVIDIA Blackwell and AMD MI400.
- Licence: Modified MIT, permissive with attribution; large deployments must credit Kimi K3.
What are the Kimi K3 open weights and what ships on 27 July?
Kimi K3 is Moonshot AI's flagship large language model, and the 27 July release publishes its complete parameter file rather than only API access. The download is roughly 594GB in MXFP4 safetensors, distributed through Moonshot's Hugging Face organisation together with a technical report and a vLLM build that includes Kimi Delta Attention (KDA) prefill cache support.
The licence is a Modified MIT variant: broadly permissive, with an attribution clause that requires large deployments to credit Kimi K3. In practical terms, you can fine-tune, quantise, redistribute derivatives, and run the model commercially without paying Moonshot, provided the attribution requirement is honoured.
That combination, near-frontier scores plus a redistributable checkpoint at this scale, is what makes the drop structurally different from Moonshot's earlier releases. If you want a walkthrough of the runtime plumbing after weights land, our companion piece on Kimi K3 setup and gotchas covers the boot path.
How does K3's architecture differ from earlier Kimi models?
K3 is a Sparse Mixture-of-Experts (MoE) model with 2.8 trillion total parameters, of which about 50 billion are active per token. The router selects 16 of 896 specialist subnetworks for each forward pass, so inference cost scales with the active slice rather than the full parameter count.
Three architectural choices matter for anyone planning to serve K3:
- Kimi Delta Attention (KDA): a decoding-side attention variant that Moonshot reports as roughly 6.3× faster on million-token contexts than the K2.6 baseline. The vLLM integration ships with a KDA-aware prefill cache.
- Stable LatentMoE: a routing and normalisation scheme intended to reduce expert collapse during long training runs.
- Attention Residuals (AttnRes): an efficiency layer Moonshot describes as delivering roughly 25% token-efficiency improvement for around 2% additional compute per step.
Context length is 1,048,576 tokens, matching the current frontier band, though real-world quality at maximum context still depends on prompt structure and retrieval design — validate at your actual depth before committing to the longest claims.
What does the independent benchmark data say?
Independent evaluation matters here because K3's own benchmark charts are the point of contention. The figures below come from Artificial Analysis's API testing rather than Moonshot's self-reported numbers.
| Metric | K3 result | Context |
|---|---|---|
| Intelligence Index | 57.1 | Rank 4 overall; behind Claude Fable 5 (59.9) and GPT-5.6 Sol Max (58.9) |
| Frontend Code Arena | 1,679 pts | Rank 1 overall, 76% win rate in blind developer voting |
| Peer band | ~ Claude Opus 4.8, ~ GPT-5.5 | On aggregate intelligence, not on coding |
| AA-Omniscience hallucination rate | ~51% | Up from 39% on K2.6 |
| AA-Omniscience accuracy | 46% | Up from 33% on K2.6 |
Two patterns are worth pulling out. First, K3's coding lead is real and third-party confirmed, not a self-reported artefact. Second, K3 got measurably better at producing correct answers and measurably worse at knowing when it should decline, which is the shape of a model that has been optimised for confidence.
For a head-to-head with the closed frontier, our Kimi K3 vs GPT-5.6 Sol comparison breaks the numbers down by task.
Why is the 51% hallucination rate the real story?
Moonshot's release materials highlight the Intelligence Index rank and the Frontend Code Arena win. They do not include the AA-Omniscience hallucination panel, even though it comes from the same evaluator whose other results Moonshot happily reproduces.
For context, Claude Fable 5 scores 54.9% on the same benchmark, so 51% is not an outlier in absolute terms. The issue is trajectory and disclosure. K3 climbed from 39% to roughly 51% while its raw accuracy also improved — the model now gives more correct answers and more confidently wrong ones, and the lab publishing the weights chose not to surface the second half.
For an open-weight release, that matters more than for a closed API. Once weights are downloadable, third parties fine-tune, quantise, and redeploy them in settings where the base honesty profile is difficult to audit. If you are building on K3 for factual workloads, the practical implication is straightforward: gate outputs behind retrieval or a verifier model, and run your own factuality suite before production.
The broader pattern, where open-weight leaders ship stronger capabilities alongside weaker disclosure than closed labs, is the shift worth tracking. Our structural analysis of the open-weight distillation debate sits alongside this piece.
What hardware and cost profile does self-hosting require?
The minimum viable configuration Moonshot describes is 8× H100 80GB, and even that is classified as experimental. Production-grade throughput starts around 64 accelerators, which puts self-hosting firmly in the datacentre bracket rather than the workstation bracket.
Native MXFP4 execution is limited to NVIDIA Blackwell and AMD MI400 hardware. On earlier GPUs the model runs, but throughput and memory pressure change materially because MXFP4 has to be upcast.
If you would rather use the API for early evaluation, Moonshot's pricing is $3.00 per million input tokens (cache-miss), $0.30 per million cached input tokens (a 90% cache discount), and $15.00 per million output tokens. Output pricing is up from K2.6's $4.00 per million, which materially changes the economics of long-response workloads.
For teams considering an open-weight model chosen for fine-tuning ergonomics rather than raw scale, Inkling's 975B MoE is a very different design point worth reading against K3.
Should you deploy K3 in production?
The honest answer depends on the workload profile.
- Frontend and full-stack code generation: K3 leads the Arena on independently verified voting. If your evaluation matches your production distribution, this is a defensible choice.
- Long-context reasoning: the million-token window and KDA make K3 competitive, but validate quality at your actual context depth rather than at the maximum.
- Factual question answering, RAG grounding, agent tool use: the AA-Omniscience trajectory is a warning. Add a verifier, ground in retrieval, or fall back to a model with a lower fabrication profile for the final answer step.
- Regulated or sensitive data: note the data-handling context. In April 2026, Kimi's hosted service disclosed one user's resume to an unrelated user during a PowerPoint translation, an incident catalogued by the OECD AI Incidents Monitor. Self-hosted deployments avoid that vector but do not avoid the licensing considerations around who can operate Blackwell hardware in which jurisdictions.
Our side-by-side model testing guide walks through the evaluation harness we use for these decisions.
FAQ
Q: When and where do the Kimi K3 open weights become available?
A: 27 July 2026 at 00:00 UTC (26 July, 8pm ET) on huggingface.co/moonshotai, alongside the technical report and a vLLM build with KDA prefill cache support.
Q: What licence covers the Kimi K3 weights? A: A Modified MIT licence. It is permissive for commercial use, fine-tuning, and redistribution, with an attribution requirement obliging large deployments to credit Kimi K3.
Q: How large is Kimi K3 and what are the active parameters? A: 2.8 trillion total parameters in a Sparse Mixture-of-Experts layout, with about 50 billion active per token. The router picks 16 of 896 experts for each forward pass.
Q: Why do independent testers report a 51% hallucination rate? A: On the AA-Omniscience benchmark, Artificial Analysis measured K3 hallucinating on roughly 51% of confident answers, up from 39% on K2.6. Moonshot omitted this panel from its published charts.
Q: What hardware do I need to self-host Kimi K3? A: A minimum of 8× H100 80GB for experimental use and 64 or more accelerators for production. Native MXFP4 execution is supported on NVIDIA Blackwell and AMD MI400 only.
Q: Is Kimi K3 stronger than closed frontier models like Claude Fable 5 or GPT-5.6 Sol Max? A: On overall Intelligence Index, no. K3 sits at 57.1 versus 59.9 and 58.9 respectively. On Frontend Code Arena it leads both, with 1,679 points and a 76% win rate in blind voting.
Discussion
0 comments