Moonshot AI Open-Sources a 1-Trillion-Parameter Coding Model at 12x Less Than GPT-5.5

By Hector Herrera | June 13, 2026 | Science · Quick Take

Moonshot AI released Kimi K2.7-Code as a fully open-source model under a Modified MIT license — a 1-trillion-parameter Mixture-of-Experts architecture priced at $0.95 per million input tokens, up to 12x cheaper than GPT-5.5 and Claude Fable 5. Weights are freely available on Hugging Face for self-hosting today.

The Numbers

• Total parameters: 1 trillion
• Active parameters per inference: 32 billion (only a fraction of total weights activate per query — that's how MoE keeps cost down)
• License: Modified MIT — open for commercial use with attribution
• Price (hosted API): $0.95 per million input tokens
• Benchmark improvement: 21.8% over Kimi K2 on Kimi Code Bench v2
• Weights: Available now on Hugging Face

What Mixture-of-Experts Means Here

MoE — Mixture-of-Experts — is an architecture where a model routes each token through a subset of specialized "expert" sub-networks rather than the full weight matrix. The result: 1 trillion parameters of capacity, but only 32 billion activate per forward pass. That keeps inference fast and cheap even at frontier scale.

The same general architecture underlies Mistral's Mixtral models and is believed to power several closed frontier models. Moonshot applying it at 1-trillion-parameter scale in an open release is notable.

Who This Is For

If you run a coding-heavy workflow — code generation, review, test writing, documentation — and your current stack uses GPT-5.5 or Fable 5, Kimi K2.7-Code is worth a cost comparison run. At 12x lower price per token, even a meaningful quality trade-off still pencils out for many high-volume applications.

For self-hosters with the hardware to run a 32B-active-parameter model (roughly equivalent VRAM requirement to other 32B-class models), the Modified MIT license means there are no restrictions on commercial deployment.

The Bigger Picture

Chinese open-source labs — Moonshot, DeepSeek, Alibaba's Qwen team — are running a sustained pressure campaign on frontier pricing. Each release forces Western providers to either match prices or justify the premium with performance. Kimi K2.7-Code's 21.8% benchmark improvement over its predecessor, combined with the price point, makes it harder to dismiss as a cost-quality trade-off.

The model launches the same week that U.S. export controls forced Anthropic to pull Fable 5 offline. Frontier capacity is becoming geographically fragmented; open-source releases with no access restrictions fill the gap.