Kimi k1.5
Kimi k1.5 is Moonshot's research model for reinforcement learning across text and vision. The public report documents training methods and evaluations, rather than a downloadable checkpoint with a verified current API.
Access not confirmed
Reuse part of a reasoning trajectory
The project scales reinforcement-learning contexts to 128K and uses partial rollouts. It can reuse a large part of an earlier trajectory rather than generate it again from the start. That is a training-efficiency method, not a published API context guarantee.Kimi k1.5 research project
Optimize reasoning without a separate search tree
The research combines policy optimization, sampling, length penalties and data selection. It reports learning useful reasoning without depending on Monte Carlo tree search, a value function or process-reward models in that framework.Kimi k1.5: original reinforcement-learning report
Long reasoning can train a shorter response model
The paper also studies long-to-short methods that transfer gains from extended reasoning into shorter responses. Its long and short evaluation settings must be kept separate; they do not establish two independently downloadable releases.Kimi k1.5: original reinforcement-learning report
Research publication is not a current service offer
Practical guidance. Use this record to understand the training contribution. For a new application, choose a model whose current access and deployment requirements are documented. Do not infer an endpoint or weight license from a research paper.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Access not confirmed |
| Launch date | January 2025 |
| Model identifiers | Kimi-k1.5 |
| Research scope | Text and visual reasoning, with long and short reasoning evaluations.Kimi k1.5: original reinforcement-learning report |
| Current access | No current endpoint or downloadable checkpoint is established by the reviewed research release.Kimi k1.5 research project |
Related models
- Kimi-K2-Thinking
Moonshot AI · Kimi K2
Long-reasoning K2 model trained to alternate internal reasoning with repeated tool interactions.
Launch date Downloadable - Kimi-Dev-72B
Moonshot AI · Kimi Dev
Qwen2.5-72B adaptation trained to locate software faults and produce repository-level fixes.
Launch date Downloadable