Beijing-based Moonshot AI has published the full weights of Kimi K3, a 2.8-trillion-parameter model that is now the largest open-weight system ever released. Launched on July 16 and shipped as downloadable weights on July 27, K3 pushes the open-source frontier well past DeepSeek's V4-Pro and lands as one of the clearest signals yet that the gap between freely available models and the best proprietary systems is narrowing fast.
A Record-Setting Architecture
Kimi K3 is a sparse Mixture-of-Experts (MoE) model — a design that keeps total parameter counts enormous while activating only a fraction of the network for any given token. The model repository describes 896 routed experts, of which sixteen fire per token alongside two shared experts, yielding roughly 104 billion active parameters against the 2.8-trillion total. It carries a one-million-token context window, accepts text, image and video input, and exposes low, high and max reasoning effort rather than a single setting.
Moonshot trained the model quantization-aware using an MXFP4 weight format, a choice that matters for anyone hoping to actually run it. The published weights ship as 96 shards totaling about 1.56 terabytes on Hugging Face, released under a custom Kimi K3 License rather than a permissive MIT-style license.
Strong, but Not Unbeatable
On the benchmark front, Moonshot has been aggressive. K3 posted the top open-weight result on the Artificial Analysis Intelligence Index, and at launch it debuted at No. 3 on that leaderboard, trailing only Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol among all models. In some practical evaluations it led outright — developers preferred it in a blind front-end coding test, and it ranked first in several task-automation benchmarks.
The company positioned K3 directly against the strongest closed models and reported wins over some of them on coding and agent tasks. But the picture deserves nuance. Moonshot's own launch materials acknowledge that K3 still trails the best proprietary systems overall, and its comparison tables mix several different agent harnesses — a reminder to read frontier benchmark claims carefully rather than at face value.
The Cost of "Open"
Open weights do not mean small or cheap to run. Moonshot recommends supernode configurations with at least 64 accelerators, and realistic self-hosting demands a multi-node cluster with well over a terabyte of aggregate GPU memory. The full model exceeds any single GPU or even a standard eight-GPU node.
The community has moved quickly to shrink that footprint. A 1-bit dynamic quantization reportedly cut the model from 1.56 TB to about 594 GB while retaining most of its accuracy — small enough, in principle, to run on a single high-memory workstation. That gap between the full release and its compressed variants captures a defining tension of the open-weight era: the weights may be free, but useful access still depends on how much hardware a team can bring, and clever quantization is fast becoming as important as the model itself. On the hosted side, official API pricing has been set at $3 per million input tokens and $15 per million output, with cached input far cheaper, matching the headline rate of some mid-tier proprietary models.
Licensing terms stay generous at ordinary scale but add conditions for the largest deployments: model-as-a-service businesses earning above a defined annual threshold need a separate agreement, and very high-traffic products must display the Kimi K3 name in their interface.
Why It Matters
Kimi K3 sharpens the central tension in frontier AI: open versus closed.
- The frontier is no longer a walled garden. A freely downloadable model now sits within striking distance of the top proprietary systems on public leaderboards.
- Openness shifts to hardware and deployment. With weights public, the real barriers become the compute and engineering needed to serve a model this large.
- China's open-model momentum is accelerating. K3 follows a string of competitive open releases from Chinese labs, reshaping who sets the pace in open AI.
For researchers and enterprises, an open model of this caliber means more control, on-premise deployment options and freedom from a single vendor — at the price of substantial infrastructure. For the broader field, K3 is another data point in a summer of compressed, overlapping frontier releases, where open and closed labs are shipping simultaneously rather than in turn. The lesson is not that open has won, but that it can no longer be dismissed. The frontier now has more than one front door.
