Kimi K3
The largest open-weight model ever released came out of Beijing, and it beat the strongest closed models on the benchmark developers cared about most.
Moonshot AI released the full weights of Kimi K3 on July 27, 2026. At 2.8 trillion parameters it was the largest openly available model yet published, distributed on Hugging Face across 96 weight shards under the Kimi K3 License along with configuration files and deployment instructions.
The benchmark results placed it firmly among the frontier rather than merely near it. On GDPval-AA v2, which measures real-world tasks across 44 occupations and nine industries, K3 scored 1,687 and ranked third overall behind only the top configurations of Claude Fable 5 and GPT-5.6. On the Artificial Analysis Intelligence Index it scored 57, ahead of Claude Opus 4.8 and behind Fable 5. Most striking, it took first place in Arena's Frontend Code evaluation at 1,679 points, ahead of every closed model in blind developer testing.
That last result was the one that travelled. Open models had been catching up on academic benchmarks for two years, but leading a blind head-to-head evaluation of practical coding work against the best-funded closed systems in the world was a different kind of claim, and it was made by a laboratory operating under United States compute export restrictions.
Key Facts
- 01Full weights released July 27, 2026, on Hugging Face across 96 shards under the Kimi K3 License.
- 02At 2.8 trillion parameters, the largest open-weight model ever published.
- 03Scored 1,687 on GDPval-AA v2, third overall behind only top configurations of Claude Fable 5 and GPT-5.6.
- 04Ranked first in Arena's Frontend Code evaluation at 1,679 points, ahead of every closed model in blind developer testing.
- 05Developed by Moonshot AI in Beijing, under United States compute export restrictions.
K3 completed an arc that DeepSeek R1 had begun eighteen months earlier. R1 showed frontier capability did not require frontier training budgets. DeepSeek V4 showed near-frontier serving costs could fall by an order of magnitude. K3 showed an open model could simply win a category outright. The question of whether open weights would remain permanently behind the closed frontier was, in at least one measurable domain, answered.
The geopolitical dimension is inseparable from the technical one. The largest open model in the world being published by a Chinese laboratory operating under export controls complicated the assumption that compute restrictions would translate cleanly into capability restrictions. It also meant that the most freely available frontier-class systems, the ones any government, company, or researcher could download and inspect without permission, were increasingly not produced in the United States. For anyone deciding what to build on, the calculation now included where the weights came from and what that implied about the future of access.
China's Moonshot AI releases Kimi K3, the largest open-source model ever, rivaling top U.S. systems
VentureBeat · 2026
https://venturebeat.com/technology/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems
Moonshot releases 2.8-trillion-parameter Kimi K3
Tom's Hardware · 2026
https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3
Kimi K3 on OpenRouter: API Pricing and Benchmarks
OpenRouter · 2026
https://openrouter.ai/moonshotai/kimi-k3