Kimi K3 Open Weights Ship 2.8T Parameters and 1M Context

The Largest Open-Weight Model Yet Is Now a File Anyone Can Download

Moonshot AI published the Kimi K3 open weights late on July 26, 2026, a day ahead of the schedule the lab had signalled, and the release is notable less for any single benchmark number than for what it makes ordinarily available. Kimi K3 is a 2.8 trillion parameter mixture-of-experts model with native vision and a one million token context window, and the weights sit on Hugging Face as roughly 1.56 TB of MXFP4 shards that any team can inspect, fine-tune, or serve themselves. Whatever else is true of the current model landscape, a frontier-class open-weight release at this scale did not exist a month ago.

• 2.8 trillion total parameters, with 104 billion active per token — a sparse mixture-of-experts design using 896 experts and 16 routed per token
• 1 million token context window, with an attention variant Moonshot calls Kimi Delta Attention that the lab reports as up to 6.3x faster decoding at million-token lengths
• 88.3% on Terminal-Bench 2.1, and a first-place finish in the Frontend Code Arena blind evaluation at 1,679 points
• Weights released under a modified MIT-style license permitting commercial use with attribution, with additional terms attached to very high-revenue deployments

What Makes Kimi K3 Different From a Dense Model?

The headline parameter count deserves a caveat that matters enormously in practice. Kimi K3 is sparse, not dense. Of the 896 experts in the network, the router activates 16 for any given token, so the compute cost per token tracks the 104 billion active parameters rather than the full 2.8 trillion. This is the same architectural bargain that has defined most large open releases over the past two years — memory capacity is expensive, but per-token compute stays tractable — and K3 pushes it further than any openly published model before it.

The second architectural note is Kimi Delta Attention, Moonshot's approach to keeping long-context decoding affordable. Standard attention costs scale unpleasantly as the context grows, and a million-token window is precisely where that pain becomes acute. Moonshot reports up to 6.3x faster decoding at the top of the context range, which is the difference between a million-token window being a specification and being something a team would actually use in a production loop.

How Does Kimi K3 Perform on Coding Benchmarks?

The evaluation picture is one of specialization rather than uniform dominance. Kimi K3 scores 88.3% on Terminal-Bench 2.1, placing it a fraction behind the strongest reported result on that benchmark. In the Frontend Code Arena blind developer evaluation it ranked first at 1,679 points. Independent aggregators placed it fourth on the Artificial Analysis Intelligence Index, with strong showings on agentic and long-horizon coding evaluations.

Third-party testing from Fireworks AI found the more useful framing: compared head-to-head against a leading closed frontier model, K3 did not win everywhere, but it was distinctly stronger in some domains — security and cryptography tasks among them. That is the practical read for anyone building on it. Treat K3 as a very strong specialist you can host yourself rather than a drop-in replacement for everything.

What Does It Take to Run Kimi K3 Locally?

Honestly, more than most readers of our AI coverage have on hand. At 1.56 TB of MXFP4 weights, K3 does not fit on a single GPU or on a standard eight-GPU node. Practical self-hosting means a multi-node cluster with roughly 1.4 TB or more of aggregate GPU memory, which in most deployments works out to 64 or more accelerators. Teams who want the model without the cluster can reach it through Moonshot's API at $3 per million input tokens — $0.30 cached — and $15 per million output tokens, and several inference providers announced day-zero hosted access timed to the drop.

For anyone weighing what a smaller quantized model can do on hardware they already own, our guide to LLM quantization formats is the more relevant starting point. K3 is a datacenter model that happens to be open, not a laptop model.

Why an Open Release at This Size Matters

The most interesting finding in early production testing is not a benchmark at all. Several teams reported that routing work between K3 and a frontier closed model produced higher accuracy than either alone, while cutting cost on long agentic workflows by a substantial margin. That is a genuinely new option in the toolbox: an openly available model strong enough to carry the bulk of a pipeline, with a premium model reserved for the passes that need it.

It also continues a run of consequential open-weight releases this month, following Thinking Machines' 975B-parameter Inkling. Weights that anyone can download are weights that anyone can audit, fine-tune for a narrow domain, or run inside their own network without sending a token outside it. At 2.8 trillion parameters, that option now extends further up the capability curve than it ever has.

Sources: Tom's Hardware — July 27, 2026; Yotta Labs — July 2026; Forbes — July 27, 2026.