Kimi K2.7 Code is the cheap open model to watch
4 reasons Kimi K2.7 Code matters: open weights, 256k context, strong coding, and $4 output pricing.
Kimi K2.7 Code is an open coding model with 256k context and very low token pricing.
Moonshot AI’s Kimi K2.7 Code is the open model that changes the cost math for coding agents: $4 output per million tokens, 256,000-token context, and a trillion-parameter MoE design.
| Item | Output price per 1M tokens | Context window | Notable benchmark |
|---|---|---|---|
| Kimi K2.7 Code | $4 | 256,000 | MCP Mark Verified 81.1 |
| GPT-5.5 | $30 | Not stated here | MCP Mark Verified 92.9 |
| Claude Opus 4.8 | $25 | Not stated here | MCP Mark Verified 76.4 |
1. A trillion-parameter coding model
Get the latest AI news in your inbox
Weekly picks of model releases, tools, and deep dives — no spam, unsubscribe anytime.
No spam. Unsubscribe at any time.
Kimi K2.7 Code is built as a Mixture-of-Experts model with 1 trillion total parameters, but only 32 billion active per token. That matters because it lets Moonshot scale the model up without making every request pay the full compute bill.
The model also uses 384 experts and a 256,000-token context window, so it can keep entire repos, long prompts, and multi-file edits in memory. Moonshot added MoonViT too, which gives it vision input for screenshots, diagrams, and mockups.
- Architecture: Mixture-of-Experts
- Total parameters: 1 trillion
- Active parameters per token: 32 billion
- Context window: 256,000 tokens
2. The strongest open model for coding
On Moonshot’s own coding tests, K2.7 Code is not the top model overall, but it is the strongest open option in the set. It beats K2.6 across every benchmark in the company’s published table, which suggests the newer release is a real step up rather than a minor refresh.
The headline is not that it outruns GPT-5.5 or Claude Opus 4.8 everywhere, because it does not. The point is that the gap is now narrow enough to matter for real projects, especially when the model is used for agentic coding and tool use instead of only raw code completion.
- Kimi Code Bench v2: 62.0
- Program Bench: 53.6
- MLS Bench Lite: 35.1
- MCP Mark Verified: 81.1
3. Pricing that changes agent budgets
This is where Kimi K2.7 Code becomes hard to ignore. Moonshot prices output at $4 per million tokens, while GPT-5.5 is listed at $30 and Claude Opus 4.8 at $25. For teams running coding agents all day, that gap can decide whether a product is viable.
The input side is cheap too at $0.95 per million tokens, and caching can drop it to $0.19. Moonshot also says K2.7 Code uses about 30% fewer reasoning tokens than K2.6, which helps reduce both latency and spend.
Pricing snapshot
- Kimi K2.7 Code: $0.95 input / $4 output
- GPT-5.5: $5 input / $30 output
- Claude Opus 4.8: $5 input / $25 output4. Open source, with a licensing catch
Moonshot publishes the weights on Hugging Face, and the model is available through the API as well. For most users, that means real open access: download it, test it, fine-tune it, or plug it into tools that already speak OpenAI-style APIs.
The catch is the modified MIT license. If a product built on K2.7 Code crosses 100 million monthly active users or $20 million in monthly revenue, it must show the Kimi K2.7 name in the interface. That still leaves startups and smaller teams with a very permissive setup.
- Weights: public on Hugging Face
- License: modified MIT
- Commercial use: allowed
- Attribution trigger: 100M MAU or $20M monthly revenue
5. How to try it without buying servers
There are three practical ways to use the model. The easiest is Kimi Code, Moonshot’s terminal agent, which starts at $19 a month. The second is the API, which is compatible with the OpenAI format and easier to slot into existing workflows.
The third is local deployment, but that is only realistic for very strong workstations or servers. Even with quantized builds such as GGUF and INT4 conversions, a trillion-parameter model still demands serious hardware.
Ways to try Kimi K2.7 Code
1. Kimi Code terminal agent
2. API with OpenAI-style format
3. Local run with quantized weightsWhat to pick
If you want the best open coding model with aggressive pricing, Kimi K2.7 Code is the obvious candidate. If you need the absolute strongest closed model on Moonshot’s published numbers, GPT-5.5 still leads several benchmarks, and Claude Opus 4.8 remains competitive on some tasks.
For most builders, the decision comes down to budget and trust. K2.7 Code is the one to test first if you care about agent costs, long context, and open weights. Just treat Moonshot’s benchmarks as company data until independent suites confirm them.