Model Details
Kimi K2 Instruct
Quick answer: Kimi K2 Instruct is Moonshot AI's July 2025 open-weight model scoring 93.3% on HumanEval, 81.1% on MMLU-Pro, 59.1% on BFCL-v3, and 9.8% on AetherCode. Under Apache 2.0 with open weights, it delivers top-tier function-calling and academic knowledge benchmarks for an open model.
At a Glance
Where Kimi K2 Instruct leads
- 93.3% HumanEval — highest among open-weight models on this benchmark
- 81.1% MMLU-Pro — matches Llama 4 Maverick (80.5%) on academic knowledge
- 59.1% BFCL-v3 — strong tool use and function calling capability
- Apache 2.0 — unrestricted commercial use
- March 2025 knowledge cutoff — more recent than most 2024-era open models
Where it lags
- 9.8% AetherCode — below Qwen3 32B (16.3%) on agentic coding tasks
- Architecture/parameters undisclosed — harder to predict inference requirements
- Moonshot AI is less established than Meta/Alibaba in the open-weight ecosystem
Best for: Open-weight deployments requiring strong function calling (59.1% BFCL-v3), high coding completion (93.3% HumanEval), and broad academic knowledge (81.1% MMLU-Pro).
What Kimi K2 Instruct Is
Kimi K2 Instruct is Moonshot AI's July 2025 instruction-tuned open-weight model. Moonshot AI, the Chinese AI lab behind the Kimi consumer product, released K2 with Apache 2.0 licensing — a notable choice that enables unrestricted commercial deployment.
The model's 93.3% HumanEval score is exceptional for an open-weight model, and its 59.1% BFCL-v3 result demonstrates strong function-calling capability — important for tool-using agent deployments. At 81.1% MMLU-Pro, it matches Llama 4 Maverick on academic knowledge benchmarks.
The 9.8% AetherCode score is lower than expected given its HumanEval performance, suggesting that Kimi K2 Instruct is strong at code completion and synthesis tasks but less effective at the end-to-end agentic software engineering measured by AetherCode.
Specifications
| Field | Value |
|---|---|
| Organization | Moonshot AI |
| Parameters | Undisclosed |
| License | Apache 2.0 |
| HuggingFace | Kimi/K2-Instruct |
| Release date | July 1, 2025 |
| Knowledge cutoff | March 2025 |
| Modality | Text only |
Pricing
Kimi K2 Instruct is available as open weights under Apache 2.0 — free to self-host. Available via Moonshot AI API and third-party providers.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| HumanEval | 93.3% | Benchgen evaluation | 2025-07 |
| MMLU-Pro | 81.1% | Benchgen evaluation | 2025-07 |
| BFCL v3 | 59.1% | Benchgen evaluation | 2025-07 |
| AetherCode | 9.8% | Benchgen evaluation | 2025-07 |
Kimi K2 Instruct vs Alternatives
| Model | HumanEval | MMLU-Pro | BFCL-v3 | License |
|---|---|---|---|---|
| Kimi K2 Instruct | 93.3% | 81.1% | 59.1% | Apache 2.0 |
| Llama 4 Maverick | — | 80.5% | — | Llama 4 |
| Phi-4 | 82.6% | 70.4% | — | Apache 2.0 |
| Qwen3 32B | — | — | — | Apache 2.0 |
Kimi K2's 93.3% HumanEval is the highest we track for any open-weight model, and its 59.1% BFCL-v3 indicates strong function-calling performance. For applications requiring agentic end-to-end coding, its 9.8% AetherCode is a limitation; Qwen3 32B (16.3%) performs better on that benchmark.
Related Models & Benchmarks
- Compare: Llama 4 Maverick · Phi-4 · Qwen3 32B
- Benchmarks: HumanEval · MMLU-Pro · BFCL v3 · AetherCode
Frequently Asked Questions
What is Kimi K2 Instruct?
Kimi K2 Instruct is Moonshot AI's July 2025 Apache 2.0 open-weight model, scoring 93.3% HumanEval, 81.1% MMLU-Pro, and 59.1% BFCL-v3. Designed for strong code completion and function calling.Is Kimi K2 Instruct open source?
Yes. Kimi K2 Instruct is released under Apache 2.0 — free for commercial use without restrictions.What is Kimi K2 Instruct strong at?
HumanEval (93.3%), MMLU-Pro (81.1%), and BFCL-v3 function calling (59.1%). It is weaker at AetherCode (9.8%) — end-to-end agentic coding tasks.Specs from Moonshot AI's official Kimi K2 release (July 2025) and Benchgen evaluations. Last updated 2026-07-24.