News · 2026-07-27

Moonshot releases Kimi K3: a 2.8-trillion-parameter open-weight model, 1.56 terabytes on disk

Moonshot AI has published the complete weights for Kimi K3, a 2.8-trillion-parameter model that anyone can download, inspect, fine-tune and run. The checkpoint is 1.56 terabytes spread across 96 files, it accepts images as well as text, and it reads a context window of 1,048,576 tokens. It is the largest frontier-class model ever released for public download, and it arrived with same-day support from two production serving engines.

Key facts

• 2.8 trillion parameters total, about 104 billion active per token - each token is routed to 16 of 896 expert networks plus two shared experts.
• Released 27 July 2026 on Hugging Face as a 96-shard checkpoint totalling 1.56 terabytes.
• Built by Moonshot AI, the Beijing lab behind the Kimi series, under a permissive but conditional commercial licence.
• Primary source: the Kimi K3 model card on Hugging Face and the accompanying technical report on GitHub.

For weeks the Kimi K3 story was a countdown rather than a release. Moonshot teased the model, benchmark screenshots circulated, and a separate embarrassment about a rival open model shipping with missing tensors made the whole open-weight scene twitchy about whether announced checkpoints would actually be complete and loadable. That question is now settled: an independent developer who converted the release reports reading and mapping all 2,760 tensors across the text, projector and vision shards with no mismatches. The weights are all there.

What is actually inside

The interesting part is not the headline size. It is the architecture, which Moonshot titled its report around: Kimi K3: Open Frontier Intelligence.

K3 is a 93-layer hybrid. Sixty-nine of those layers use what Moonshot calls Kimi Delta Attention, a recurrent, fixed-size memory that does not grow as the conversation gets longer. The other 24 are conventional global attention layers. The pattern runs roughly three recurrent layers to every global one, ending on a global layer.

The analogy is a note-taker in a long meeting. Ordinary attention keeps a verbatim transcript of everything said, which is perfect recall but grows without limit - this is the KV cache problem that makes long chats expensive. K3's recurrent layers instead keep a running summary of fixed size, and every fourth layer stops to re-read the actual transcript. You get most of the fidelity of full attention at a fraction of the memory cost, which is how a million-token window becomes affordable rather than theoretical. Readers who want the underlying idea can start with our explainer on linear attention.

Two other pieces are genuinely new. Attention Residuals lets a layer attend backwards across the model's own depth, retrieving useful representations from earlier blocks instead of simply summing them. Stable LatentMoE projects into a narrower latent space before routing to experts, cutting the bandwidth cost of moving activations around, and uses a technique Moonshot calls Quantile Balancing to keep all 896 experts evenly used without hand-tuned load-balancing constants. If mixture of experts is new to you, that lesson covers the basic committee-of-specialists idea this builds on.

The model was also trained natively in MXFP4 weights with MXFP8 activations, rather than being trained at full precision and squeezed afterwards - a meaningful difference for anyone who has watched a quantized model lose quality on the way down.

Open weights, with fine print

The licence permits use, modification, redistribution, fine-tuning and sale. But a model-as-a-service business above roughly 20 million dollars in aggregate trailing-twelve-month revenue needs a separate agreement with Moonshot for commercial use, and sufficiently large commercial products must visibly credit Kimi K3. That is broadly permissive, and it is not MIT or Apache. The distinction matters for the policy fight now running through the industry, which our lesson on open versus closed models unpacks in plain terms.

How good is it, really

Independent evaluation is encouraging without being a coronation. Artificial Analysis headlined its writeup "Kimi K3 achieves #3 in the Artificial Analysis Intelligence Index", placing it comparable to Opus 4.8 and GPT-5.5 and behind Claude Fable 5 and GPT-5.6 Sol, while noting the first-party API is slow and unusually verbose. The strongest early counter-take comes from the independent Prinzbench author, who rates K3 their best listed open model but describes it as inconsistent - failing some easy legal-research questions while solving harder ones.

Moonshot's own benchmark table is unusually candid about differing harnesses, tool settings and competitor fallbacks, which is more honest than most launch tables and also means it is not an apples-to-apples league table. Hugging Face's newly surfaced "community evaluations" for the model were automatically extracted from that same model card, so they are not independent replications either.

Why it matters

An open frontier checkpoint changes who gets to compete on serving. Cloud providers, sovereign deployments and data-residency-conscious enterprises can now run a top-tier model on their own hardware and compete on caching, quantization and routing rather than on API access. That is the real unlock, and the Hacker News discussion reflected it: the dominant thread was not "can it beat Fable" but deployment economics.

The honest caveat

"Downloadable" is not "runnable at home." The 1.56-terabyte checkpoint and the interconnect needed to hold 896 experts resident across a GPU domain are now the binding constraints, and Moonshot's own guidance recommends supernodes of 64 or more accelerators for efficient deployment. What K3 democratizes first is the right to operate a rack. We cover that hardware floor in a separate story.

Primary source, verified: read the paper →

Ground Truth. (2026, July 27). Moonshot releases Kimi K3: a 2.8-trillion-parameter open-weight model, 1.56 terabytes on disk. Ground Truth. https://groundtruth.day/news/moonshot-ships-kimi-k3-weights-2-8-trillion-parameters.html

@misc{groundtruth:moonshot-ships-kimi-k3-weights-2-8-trillion-parameters,
  title  = {Moonshot releases Kimi K3: a 2.8-trillion-parameter open-weight model, 1.56 terabytes on disk},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/moonshot-ships-kimi-k3-weights-2-8-trillion-parameters.html}
}

Topics: open-weights · kimi-k3 · moonshot · mixture-of-experts · model-release · china