Kimi K2.7 Code Review: The First Open-Weight Model in GitHub Copilot
An Open-Weight Model Just Entered GitHub Copilot
On July 1st, GitHub dropped a quiet announcement: Kimi K2.7 Code is now generally available in GitHub Copilot. It’s the first open-weight model to appear in the Copilot model picker.
Until now, Copilot only offered closed-source models — GPT, Claude, Gemini. Now a model from Moonshot AI, a Chinese AI lab, sits right next to them. You can select it in VS Code 1.127+, Visual Studio 17.14.6+, JetBrains 1.9.1+, Copilot CLI, and even GitHub Mobile.
For enterprise users, Copilot Business and Enterprise plans have it disabled by default — administrators must explicitly enable the Kimi K2.7 Code policy. GitHub recommends reviewing it against your security and compliance requirements first. Standard language, but the fact that it’s there at all says something about the model’s quality.
From K2.6 to K2.7 Code: What Changed
Kimi K2.7 Code is built on K2.6 but serves a different purpose — K2.6 is a general-purpose model, K2.7 Code is coding-specialized.
Benchmark data from the official page and HuggingFace model card:
| Benchmark | K2.6 | K2.7 Code | Improvement |
|---|---|---|---|
| Kimi Code Bench v2 | 50.9 | 62.0 | +21.8% |
| Program Bench | 48.3 | 53.6 | +11.0% |
| MLS Bench Lite | 26.7 | 35.1 | +31.5% |
| Kimi Claw 24/7 Bench | 42.9 | 46.9 | +9.3% |
| MCP Atlas | 69.4 | 76.0 | +9.5% |
| MCP Mark Verified | 72.8 | 81.1 | +11.4% |
Two things worth noting:
- Coding gains are much larger than agentic gains. MLS Bench Lite jumped 31.5% — this benchmark asks models to invent generalizable ML methods, so K2.7 Code improved most on complex research-oriented coding tasks.
- Thinking-token usage dropped ~30%. The model reaches answers with less internal deliberation, which directly impacts response speed and API costs.
How It Stacks Up Against Top Closed-Source Models
The honest comparison (from the same official benchmarks):
| Benchmark | K2.7 Code | GPT-5.5 | Claude Opus 4.8 |
|---|---|---|---|
| Kimi Code Bench v2 | 62.0 | 69.0 | 67.4 |
| Program Bench | 53.6 | 69.1 | 63.8 |
| MLS Bench Lite | 35.1 | 35.5 | 42.8 |
| MCP Atlas | 76.0 | 79.4 | 81.3 |
| MCP Mark Verified | 81.1 | 92.9 | 76.4 |
There’s a real gap. On Program Bench — which asks models to rebuild program behavior from compiled binaries — GPT-5.5 scores 69.1 versus K2.7 Code’s 53.6. That’s significant.
But look at the other side:
- On MLS Bench Lite, K2.7 Code (35.1) is virtually tied with GPT-5.5 (35.5)
- On MCP Mark Verified, K2.7 Code (81.1) beats Claude Opus 4.8 (76.4)
- K2.7 Code is open-source. GPT-5.5 and Opus 4.8 are not.
An open-weight model matching or beating the most expensive closed-source models on some benchmarks — that’s a meaningful signal regardless of where you stand on the open vs closed debate.
1T Parameters, 32B Active — Architecture Details
| Parameter | Value |
|---|---|
| Architecture | Mixture-of-Experts (MoE) |
| Total Parameters | 1T |
| Activated Parameters per Token | 32B |
| Layers | 61 |
| Experts | 384 |
| Experts per Token | 8 |
| Context Length | 256K |
| Attention | MLA (Multi-head Latent Attention) |
| Vision Encoder | MoonViT (400M params) |
384 experts, 8 active per token — classic MoE tradeoff of capacity versus compute. The MLA attention mechanism (originally from DeepSeek V2) keeps KV cache memory in check.
The MoonViT vision encoder is worth mentioning — K2.7 Code accepts text, image, and video input. You can screenshot a UI design and have it write the frontend code, or photograph a whiteboard architecture diagram for analysis.
Pricing: Open-Source But Not Free
Via the Kimi API:
| Model | Input (cache miss) | Input (cache hit) | Output | Context |
|---|---|---|---|---|
| kimi-k2.7-code | $0.95/M | $0.19/M | $4.00/M | 262,144 tokens |
Compare that to Claude Sonnet 4.6 at $3.00/$15.00 — input is 3x cheaper, output nearly 4x cheaper. The API supports automatic context caching, so repeated context drops to $0.19/M. Long multi-turn sessions get significantly cheaper.
Since the weights are open (Modified MIT License), you can also self-host with vLLM, SGLang, or KTransformers. A 1T-parameter model isn’t trivial to deploy, but native INT4 quantization is available, and the community has Ollama builds and various quantized versions ready.
Kimi Code subscription plans (annual billing, monthly price):
| Plan | Price | Best for |
|---|---|---|
| Moderato | $15/mo | Regular coding, weekly refreshed quota |
| Allegretto | $31/mo | Larger quota + higher concurrency |
| Allegro | $79/mo | Intensive development, complex projects |
| Vivace | $159/mo | Maximum quota, large codebases |
Practical Gotchas
A few things to watch out for:
- Thinking mode is mandatory. K2.7 Code doesn’t support non-thinking mode. If you disable thinking in Kimi Code, requests automatically fall back to K2.6.
- It’s a coding model, period. For writing, analysis, or general conversation, Moonshot AI recommends K2.6 instead.
- Temperature and sampling. The recommended settings are temperature=1.0 and top-p=0.95 — higher than most models’ defaults. Adjust accordingly.
Who Should Try K2.7 Code
| Scenario | Recommendation | Why |
|---|---|---|
| GitHub Copilot users | Try it | Select it directly in the model picker, zero migration cost |
| Cost-conscious teams | Recommended | API pricing is ~1/3 to 1/4 of Claude |
| Bilingual (CN/EN) projects | Strongly recommended | Chinese coding comprehension is a genuine strength |
| Self-hosting requirements | Worth evaluating | Open weights, Modified MIT license |
| Maximum coding performance | Not yet | Program Bench and other hard benchmarks still trail GPT-5.5 |
| Non-coding tasks | Not recommended | Moonshot AI explicitly says use K2.6 for general-purpose work |
The Bottom Line
Kimi K2.7 Code isn’t just another open-source model release. It’s the first open-weight model in GitHub Copilot’s model picker. It beats Claude Opus 4.8 on MCP Mark Verified. It costs roughly a third of mainstream closed-source models. And it uses 30% fewer thinking tokens than its predecessor.
The weaknesses are clear too — it trails GPT-5.5 on the hardest coding benchmarks, forced thinking mode limits flexibility, and self-hosting a 1T model still requires serious infrastructure.
But for most real-world development workflows, K2.7 Code offers something that didn’t exist before: a good enough, cheap enough, open-source coding model that sits right in your existing tools. That changes the competitive landscape.