Morning Edition · Friday, July 17, 2026Published at 1:29 AM EDT · New York

Moonshot Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model It Says Rivals US Frontier

The mixture-of-experts model ships with a 1-million-token context and posts 93.5% on GPQA Diamond, with full weights promised by July 27.

Save

Moonshot Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model It Says Rivals US Frontier

Moonshot AI launched Kimi K3 on July 16, a mixture-of-experts model with roughly 2.8 trillion total parameters that routes each token through 16 of 896 experts, per the company's technical blog. The application programming interface (API) is live now, and full downloadable weights are slated for July 27, which if it holds would make K3 among the larger openly released models so far. Documentation seen by the AI ML Big Data channel lists a context window of up to 1 million tokens and a focus on coding, 3D and knowledge tasks.

On vendor-reported benchmarks, K3 scores 93.5% on GPQA Diamond and 88.3% on Terminal-Bench 2.1, with agentic results of 91.2% on BrowseComp and 84.2% on MCP Atlas. The AI Post channel places K3 ahead of Anthropic's Opus 4.8 and just below GPT-5.6 and Fable 5, which would narrow the historical gap between Chinese open-weight models and US closed systems.

Every figure here comes from Moonshot's own launch materials and has not been independently reproduced. The real test is whether third-party harnesses confirm the agentic and long-context scores once the weights ship, and whether serving a 2.8-trillion-parameter model at 1 million tokens of context is economical outside a well-provisioned cluster.

Veracity: Plausible

73/100

If true, who benefits: Moonshot AI and China's open-weight ecosystem gain standing and soft power from narrowing the US frontier gap, while self-hosting enterprises and sovereignty-minded governments gain a US-independent stack that pressures closed US labs' pricing power.
The nuance: Every headline benchmark is Moonshot's own and unreproduced at launch, the full weights are only promised for July 27, and early independent evaluation places K3 near Opus 4.8 but behind GPT-5.6 Sol and Fable 5, short of clearly rivaling the top US systems.

An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.

What this means

An openly downloadable model claiming near-frontier agentic and coding scores hands sovereignty-minded governments and cost-sensitive enterprises a US-independent stack they can self-host, which erodes the distribution advantage of closed labs through the weights-download channel rather than through API competition. Moonshot and Chinese open-weight vendors gain standard-setting leverage, while the exposed parties are closed labs whose pricing power depends on capability exclusivity.

What to watch

Whether the July 27 weight release actually lands on schedule and reproduces the launch benchmarks under independent evaluation.
How quickly inference providers offer hosted K3 endpoints, which would signal real deployment demand rather than a benchmark headline.

Observations to monitor, not financial advice.

2 sources

Synthesized from: Polylog editors · Kimi (Moonshot AI)

Part of a tracked trend

Open-Weight Models Close the Gap With Closed Frontier Labs

Over the next 3-9 months, open-weight releases with downloadable weights, long context, and strong agentic/coding performance increasingly match closed frontier models on practical work, eroding the closed-lab moat.