Moonshot AI released Kimi K2.7 Code on June 12, 2026, and it immediately turned heads. A trillion-parameter open-weight model designed specifically for agentic coding tasks, with a Modified MIT license and pricing that undercuts most closed alternatives. Here's what matters and where it fits in the June 2026 AI landscape.
What Is Kimi K2.7 Code?
Kimi K2.7 Code is a Mixture-of-Experts (MoE) transformer built for long-horizon software engineering. The architecture keeps a massive parameter count accessible by only activating a fraction per token:
| Spec | Detail |
|---|---|
| Total parameters | 1 trillion (~1.1T on disk) |
| Active parameters per token | 32B |
| Experts | 384 (8 selected + 1 shared) |
| Layers | 61 |
| Attention type | MLA (Multi-head Latent Attention) |
| Context window | 256K tokens |
| Vision support | MoonViT encoder (image + video) |
| License | Modified MIT (open weights) |
The model forces "thinking" mode on permanently. It always reasons before answering and preserves its full reasoning chain across multi-turn conversations. This is deliberate — Moonshot found that retained context improves performance in coding-agent scenarios where work builds up over many steps.
What Changed From K2.6?
Two improvements define this release:
1. Better coding benchmarks across the board. Moonshot's own benchmarks show consistent gains:
| Benchmark | K2.6 | K2.7 Code |
|---|---|---|
| Kimi Code Bench v2 | 50.9 | 62.0 |
| Program Bench | 48.3 | 53.6 |
| MLS Bench Lite | 26.7 | 35.1 |
| MCP Atlas | 69.4 | 76.0 |
| MCP Mark Verified | 72.8 | 81.1 |
The biggest jump is Kimi Code Bench v2 (+11.1 points). MCP benchmarks — which measure tool-calling and Model Context Protocol workflows — show strong gains consistent with the agentic coding focus.
2. ~30% fewer thinking tokens. Since thinking is always on and reasoning tokens are billed as output, this is a direct cost reduction on every API call, not just a quality improvement.
Pricing and Access
- API: $0.95 per million input tokens, $4.00 per million output tokens
- Model ID:
kimi-k2.7-code - Open weights: Available on Hugging Face (
moonshotai/Kimi-K2.7-Code) - Recommended inference engines: vLLM, SGLang, KTransformers
At $0.95/$4.00, it's competitive with models like DeepSeek V4-Pro for coding workloads while offering a substantially larger parameter pool.
How It Compares in June 2026
The June AI landscape is crowded. Here's where Kimi K2.7 Code sits:
vs. Claude Opus 4.8 (Anthropic, May 28): Opus 4.8 leads the Artificial Analysis Intelligence Index at 61.4 and dominates SWE-Bench Pro (69.2%). It's the stronger coding model overall — but it's closed, expensive ($5/$25), and has no open-weight option. Kimi K2.7 Code gives you open weights at a fraction of the cost.
vs. DeepSeek V4-Pro: Both are open-weight MoE models. DeepSeek holds the download lead on HuggingFace (5.8M+), but Kimi K2.7 Code's 256K context and forced thinking mode make it more tailored to agentic coding workflows.
vs. GLM-5.2 (Zhipu AI): GLM-5.2 is lighter and faster but doesn't match Kimi's raw scale or context window for complex multi-step engineering tasks.
The Honest Caveats
Every benchmark published for K2.7 so far comes from Moonshot's own proprietary suites — Kimi Code Bench, Program Bench, MCP Atlas. No independent third-party validation (SWE-bench, GPQA, AIME) was available at launch. The numbers are impressive but self-reported. Treat them as directional until independent evaluations land.
Additionally, the forced thinking mode means you can't optimize for speed by disabling reasoning. Every call pays the thinking-token cost. The 30% reduction from K2.6 helps, but if you need instant responses for simple queries, this isn't the right model.
Who Should Use It?
- Teams building coding agents who want open weights and don't want to depend on Anthropic or OpenAI
- Self-hosters with the GPU capacity to run a 1T MoE (you need serious hardware — 32B active per token is still substantial)
- Cost-conscious developers who need strong coding performance at $0.95/$4.00 per million tokens
Bottom Line
Kimi K2.7 Code is a serious entry in the open-weight coding model space. The combination of trillion-parameter scale, 256K context, forced agentic reasoning, and Modified MIT licensing makes it one of the most capable open models available right now. The self-reported benchmarks need independent validation, but the architecture and pricing make it worth testing for any team doing serious coding work.
Want to compare more models? Check out the Q4KM AI Model Directory for detailed specs, benchmarks, and comparisons across hundreds of AI models.