Moonshot AI drops a 1T-parameter coding beast

Kimi K2.7 Code is an open-source, 1-trillion parameter Mixture-of-Experts model built by Moonshot AI specifically for end-to-end software engineering and agentic workflows. It is designed to tackle complex multi-file refactoring, long debugging sessions, and autonomous coding tasks while drastically cutting down on inference bloat.

If you want an open model that can genuinely handle long-horizon coding without losing the plot halfway through, this release is worth your attention.

Trillion-parameter MoE with 32B active weights

Under the hood, K2.7 Code uses an MoE architecture with 1 trillion total parameters, routing only 32 billion active parameters per token. It packs a 256K context window, utilizes Multi-head Latent Attention, and integrates MoonViT, a 400M-parameter vision encoder.

The full model weights are openly accessible on Hugging Face. The catch is scale: the checkpoint weighs in at nearly 600GB, meaning you will need serious datacenter iron if you plan to host it yourself.

The most practical technical tweak is its restraint. Reasoning models frequently overthink simple tasks, burning through thousands of deliberative tokens. K2.7 Code cuts thinking-token consumption by roughly 30% compared to K2.6, delivering snappier responses and saving valuable context room.

Closing in on frontier closed-source models

Performance-wise, the model shows sharp jumps over K2.6 across the board. It scores +21.8% higher on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite. On agentic evaluations like Kimi Claw 24/7 Bench and MCP Atlas, it delivers a roughly 10% uplift, putting it in striking distance of proprietary heavyweights like GPT-5.5 and Opus 4.8.

You can pull the weights from Hugging Face or access it directly via Kimi Code in your browser, terminal CLI, and IDE extensions. For production pipelines, the Kimi API offers token-based billing with automatic context caching, priced at $0.19 per million cached input tokens and $0.95 for cache misses.

Keep in mind that K2.7 Code is thinking-only. You cannot toggle thinking mode off; disabling it in Kimi Code simply reverts your request back to K2.6. For non-coding tasks like writing or general chat, Moonshot AI still recommends sticking with K2.6.

Why it matters

For enterprise engineering teams that require full on-prem code privacy, K2.7 Code offers a rare open-weights alternative that genuinely rivals top-tier closed models in complex software development.

We think the smart move here is prioritizing reasoning efficiency over blind reasoning length. Trimming 30% off thinking tokens while improving long-horizon completion rates makes agentic code generation economically viable at scale. Unless you have racks of enterprise GPUs lying around, running the nearly 600GB weights locally won't be practical for individual hackers, but using it through the API or CLI gives developers an extremely sharp tool right now.