Image: Tom's Hardware / Future Moonshot AI Dropped the Largest Open-Weight Model Ever. It Beats Fable 5 on Coding.
Kimi K3 has 2.8 trillion total parameters, a 1M-token context window, and native vision. It launched July 16 on the Kimi app and API. Open weights arrive by July 27.
Moonshot AI launched Kimi K3 on July 16 with a claim that will get attention in model benchmarking circles: it beat Anthropic’s Claude Fable 5 on the Frontend Code Arena, the coding benchmark that uses blind developer votes.
The model has 2.8 trillion total parameters, a 1-million-token context window, and native vision built in. Moonshot calls it the first “open 3T-class model.” Open weights are coming by July 27.
The numbers
K3 is a sparse mixture-of-experts model. Like DeepSeek’s architecture, it keeps only a fraction of parameters active during inference, which means it runs significantly cheaper than a dense 2.8T model would. Moonshot hasn’t specified how many parameters activate per token.
On coding benchmarks:
- Frontend Code Arena: 1,679 points, first place — ahead of Fable 5 in blind developer testing
- DeepSWE: 67.5
- ProgramBench: 77.8 raw pass rate
- Terminal-Bench 2.1: 88.3
- FrontierSWE: 81.2
Moonshot’s own summary: K3 still trails Fable 5 and GPT 5.6 Sol on overall capability, but beats everything else it tested against. That includes Claude Opus 4.8 and GPT 5.5 across coding and agentic tasks.
Why it matters for open-source AI
The previous largest open-weight model was DeepSeek V4 Pro at roughly 1.6 trillion parameters. K3 doubles that. When the weights ship on July 27, anyone with the hardware will be able to run, fine-tune, or audit a model that competes on coding benchmarks with Anthropic and OpenAI’s commercial flagships.
The pricing on the API — $3 per million input tokens, $15 per million output — puts K3 in the same tier as Anthropic’s mid-range Claude models. That’s the highest price point any Chinese AI lab has posted for a public model. Moonshot is betting the performance justifies it.
Timing
The K3 release came hours before Google was expected to launch Gemini 3.5 Pro on July 17. That launch didn’t happen on schedule; Google appears to have delayed it again. The result was that every Gemini benchmark comparison this week will get stacked against a Chinese model whose weights will be free in ten days.
Kimi K3 is available now on the Kimi app and the Moonshot API. Kimi Code gets the model on launch day.
Sources: Tom’s Hardware — CNBC — Simon Willison