Kimi K2.5 is Moonshot AI's trillion-parameter multimodal agent model, released January 27, 2026. Built by continual-pretraining on ~15 trillion mixed visual and text tokens on top of the Kimi-K2 base, it natively understands text, images, and video, activates 32B parameters per token, and ships both instant and thinking modes.
Its defining feature is agentic: K2.5 introduced Agent Swarm, coordinating up to 100 specialized agents on a single task, and posted 76.8% on SWE-bench Verified — frontier-class coding from an open-weight release. The Modified MIT license is effectively free for everyone below 100M monthly users or $20M monthly revenue, at which point attribution (not payment) kicks in.
At $0.60 per million input tokens and $2.50 per million output on the first-party API, K2.5 undercuts proprietary flagships by a wide margin while being one of very few open-weight models with native video understanding.
Kimi K2.5 specifications
| Vendor | Moonshot AI (China) |
|---|---|
| Architecture | Mixture-of-Experts, natively multimodal (continual pretraining on ~15T mixed visual and text tokens) |
| Total parameters | 1T |
| Active parameters | 32B |
| Context window | 256K tokens |
| Modalities | text, image, video |
| License | Modified MIT — free commercial use; attribution required above 100M monthly active users or $20M monthly revenue |
| Release date | 2026-01-27 |
| Weights | Hugging Face · GitHub |
Kimi K2.5 API pricing
| Direction | Price per 1M tokens |
|---|---|
| Input | $0.60 |
| Output | $2.50 |
First-party Moonshot API rate as of July 2026. Verified July 2026.
Benchmarks and reported results
| Benchmark | Result | Note |
|---|---|---|
| SWE-bench Verified | 76.8% |
Running Kimi K2.5 locally
1T total parameters puts full-precision self-hosting in multi-node territory; the community INT4 builds need roughly 600GB. Realistic self-hosting means H200-class nodes running vLLM, SGLang, or KTransformers — not a desktop. Most teams use the API or a hosted provider.
What Kimi K2.5 is best for
- Multimodal agents that need vision and video, not just text
- Frontier coding (76.8% SWE-bench Verified) on open weights
- Multi-agent orchestration workloads (Agent Swarm)
Frequently asked questions
Is Kimi K2.5 open source?
Kimi K2.5 is open-weight under a Modified MIT license: code and weights are free to use commercially, with an attribution requirement that only applies above 100M monthly active users or $20M monthly revenue. Training data is not published.
How much does the Kimi K2.5 API cost?
As of July 2026, $0.60 per million input tokens and $2.50 per million output tokens on Moonshot's first-party API.
Can I run Kimi K2.5 locally?
Only on serious hardware: it is a 1T-parameter MoE, so INT4 builds still need roughly 600GB of memory across H200-class nodes running vLLM, SGLang, or KTransformers. For most teams the API or a hosted provider is the practical route.
What is Agent Swarm in Kimi K2.5?
Moonshot's multi-agent orchestration capability: K2.5 can coordinate up to 100 specialized agents working on a task simultaneously, which is how it targets long-horizon agentic workloads rather than single-shot chat.