Moonshot AI Launches Kimi K3, the World's Largest Open-Source Model at Half the Price of GPT-5.6 Sol

Moonshot AI announced on July 17 the Kimi K3, a model with 2.8 trillion parameters that debuted in first place in LMArena's Frontend Code Arena benchmark with a score of 1,679 Elo points, surpassing Anthropic's Claude Fable 5 in the interface development segment. Alex Liu, an analyst at Bank of America, sent a note to clients distributed by CNBC: "K3 raises the ceiling of capacity for AI models in China, shifting the burden of proof onto other independent labs."

In the overall ranking of Artificial Analysis, the K3 ranks third with an intelligence index of 57.11, behind Claude Fable 5 and OpenAI's GPT-5.6 Sol. The gap remains, but price is Moonshot's central argument: $3.00 per million input tokens and $15.00 per million output tokens, matching the Claude Sonnet line and half the cost of GPT-5.6 Sol.

Architecture Designed for Agents

K3 is a Mixture-of-Experts model with 896 experts, of which only 16 are activated per token, resulting in an activation rate of 1.8%. The KDA (Kimi Delta Attention) architecture, a hybrid linear attention developed in-house by Moonshot, delivers a context window of 1 million tokens with controlled inference costs. Two commercial profiles have been available since launch: K3 Max for chat and autonomous agents, and K3 Swarm Max for large-scale parallel processing. The full weights will be released by July 27, enabling local deployments without reliance on the Moonshot API. In the coding index of Artificial Analysis, K3 achieved a score of 76.24, surpassing Claude Opus 4.8 and GPT 5.5.

The Silicon That Should Not Exist

Moonshot did not disclose the hardware used to train the K3. The startup is partnered with Huawei, which this year introduced the Atlas 950 SuperPoD as a computing platform for national-scale AI. The internal compiler MiniTriton, developed by Moonshot for K3, was compared against Nvidia's Triton on an L20 GPU, a cut version of the Ada commercially sold in China under current U.S. export rules.

In January 2026, the U.S. Congress closed loopholes that allowed Chinese companies to remotely access restricted accelerators in data centers outside the country. The K3 was trained and launched after this closure. This does not demonstrate that the restrictions failed: training a model with 2.8 trillion parameters on domestic hardware involves substantially higher computational costs than on H100 or A100 clusters. The restrictions raise the marginal cost of development but do not create an insurmountable ceiling for sufficiently funded startups.

Reading in Three Markets

In India, TCS, Infosys, and Wipro are building AI practices on APIs from Western models and charging global clients for integration and compliance layers. The K3 with weights released on July 27 represents a local deployment alternative that cuts dependence on external licenses. For projects where the client processes sensitive data and wishes to avoid routing tokens through third-party APIs, a comparably performing open-weight model shifts the cost and sovereignty calculus of projects.

Using a model from a Chinese startup without disclosing hardware, data custody chain, or training audit creates direct friction under the EU's AI Act, which requires transparency about the origins and procedures of models classified as high-risk. The release of the weights on July 27 does not resolve this issue: auditing 2.8 trillion parameters for regulatory compliance is beyond the operational reach of any corporate compliance team.

At $15.00 per million output tokens and open weights for local deployment starting July 27, Moonshot is directly entering the segment where Anthropic and OpenAI compete for medium-sized corporate contracts. For American labs, K3 confirms that the competition for frontier models will not be resolved by unilateral industrial policy: development costs are rising, but China's capacity to achieve parity with Western leaders is demonstrated.