Kimi K3: The Open Model Closing the Gap

Moonshot published Kimi K3's full 2.8-trillion-parameter weights on July 27, 2026, eleven days after launch. The Hugging Face repo now holds 96 shards, the Kimi K3 License, and the technical report.

That closes the caveat in our launch-day review. K3 is no longer an open-weight promise with a date attached. The new caveat is physical: Moonshot recommends at least 64 accelerators. Open does not mean small.

The weights are real, and they are large

The official model repository describes Kimi K3 as a Mixture-of-Experts model with 2.8 trillion total parameters and 104 billion active parameters. Sixteen of 896 routed experts activate for each token, alongside two shared experts. The context window is exactly 1,048,576 tokens, and the model accepts text and images.

The release uses quantization-aware MXFP4 weights with MXFP8 activations. Moonshot lists vLLM, SGLang, and TokenSpeed as supported inference paths. Those are useful implementation receipts, but they do not turn a 2.8T model into a workstation download. The launch post recommends 64 or more accelerators because communication and expert routing become part of the serving problem at this scale.

For most teams, the direct API remains the practical way to test the model. The Kimi API pricing hub tracks $3.00 per million cache-miss input tokens, $0.30 per million cache-hit input tokens, and $15.00 per million output tokens. The rate stays flat across the full context window.

The table cannot prove a best-model claim

Publishing the weights does not settle whether K3 is the best model. Moonshot's own launch post says its overall results still trail the strongest proprietary systems. That admission is a more useful starting point than turning one winning row into a universal verdict.

The provider table is broad, but its footnotes matter. Kimi K3 runs through Kimi Code on several coding tests, while comparison rows can use Claude Code, Codex, or another published harness. Some results come from official leaderboards; others are Moonshot runs on modified hardware or internal benchmarks. Reasoning effort is set to max. A table can contain exact numbers and still compare more than one experimental setup.

We keep those distinctions on the live Kimi K3 model profile. Provider-exact rows, benchmark-native rows, display-only evidence, and ranking-eligible inputs do not collapse into one bucket. The profile is the owner for current scores; this release review should not freeze them in July prose.

That boundary changes the verdict. K3 has enough sourced evidence to be compared seriously, and the full weights make independent reproduction possible. Neither fact grants it a permanent rank. The next useful result is a run whose model, reasoning effort, harness, hardware, and task version all match.

Each Kimi path has a failure mode

Three distinct ways to use the family now exist:

Start at the Kimi pricing owner, then check Chinese model rankings or run a direct model comparison. If self-hosting is the reason for choosing K3, put the workload into the self-host calculator before treating the API price as the expensive option. GPUs have invoices too; they merely arrive under different nouns.

The weights release removes uncertainty about availability. It does not remove the work of serving, evaluating, or constraining the model. The decision to watch now is whether independent deployments can reproduce the flagship results at a cost that beats the API for a real workload.