Model Details
Kimi K2 Base
Quick answer: Kimi K2 Base is Moonshot AI's July 2025 base model scoring 92.1% GSM8K, 87.8% MMLU, 69.2% MMLU-Pro, and 70.2% MATH. Apache 2.0 — the pre-instruct base for the Kimi K2 family.
At a Glance
Where Kimi K2 Base leads
- 92.1% GSM8K — strong base-model math
- 87.8% MMLU — excellent knowledge at base level
- Apache 2.0 — fully open
- Starting point for fine-tuning the Kimi K2 architecture
Where it lags
- Not instruction-tuned: requires fine-tuning or prompting for task-specific use
- 35.3% SimpleQA — factual accuracy gap
- Use Kimi K2 Instruct variants for direct deployment
Best for: Fine-tuning and custom instruction-tuning of the Kimi K2 architecture; research on Kimi K2 pretraining; base-model evaluation studies.
What Kimi K2 Base Is
Kimi K2 Base is the pretrained base model from Moonshot AI's Kimi K2 series, released July 2025. It is the foundation for the Kimi K2 Instruct, K2 0905, and K2 Thinking variants. For direct deployment, use the instruct variant.
Specifications
| Field | Value |
|---|---|
| Organization | Moonshot AI |
| License | Apache 2.0 |
| HuggingFace | moonshotai/Kimi-K2-Base |
| Release date | July 2025 |
| Architecture | MoE |
| Modality | Text only |
Pricing
Open weights under Apache 2.0 — self-host at no cost.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| GSM8K | 92.1% | Benchgen evaluation | 2025-07 |
| MMLU | 87.8% | Benchgen evaluation | 2025-07 |
| MATH | 70.2% | Benchgen evaluation | 2025-07 |
| MMLU-Pro | 69.2% | Benchgen evaluation | 2025-07 |
| SimpleQA | 35.3% | Benchgen evaluation | 2025-07 |
Kimi K2 Base vs Alternatives
| Model | MMLU | GSM8K | Type | License |
|---|---|---|---|---|
| Kimi K2 Base | 87.8% | 92.1% | Base | Apache 2.0 |
| Kimi K2 0905 | — | 95.0% | Instruct | Apache 2.0 |
| Kimi K2 Instruct 0905 | — | — | Instruct | Apache 2.0 |
Use the instruct variants for chat/task deployment. Kimi K2 Base is for fine-tuning on top of Moonshot's pretraining.
Related Models & Benchmarks
- Compare: Kimi K2 0905 · Kimi K2 Instruct 0905 · Kimi K2 Thinking 0905
- Benchmarks: GSM8K · MMLU · MATH
Frequently Asked Questions
What is Kimi K2 Base?
Moonshot AI's July 2025 pretrained base MoE model: 92.1% GSM8K, 87.8% MMLU, 69.2% MMLU-Pro. Apache 2.0 — use instruct variants for direct deployment.Specs from Moonshot AI's Kimi K2 Base release (July 2025) and Benchgen evaluations. Last updated 2026-07-24.