Kimi K2.5
Released 2026-01 · hybrid · 262K tokens · 7 benchmarks · Open weight
Editorial notes
Lanzado 27 enero 2026. Open-weight MoE 1T totales / 32B activos. Multimodal vision + agentic. Feature 'Agent Swarm' (100 agentes coordinados). Considerado el modelo chino open-source mas fuerte. Compite directo con GPT-5.4 y Claude Sonnet 4.5.
Spec sheet
- Company
- Moonshot AI
- Country
- CN
- Type
- hybrid
- Release
- 2026-01
- Context
- 262K tokens
- Params total
- 1.0T
- Params active (MoE)
- 32B
- License
- mit-modified
- Quants
- BF16, Q8_0, Q5_K_M, Q4_K_M
- Slug
- kimi-k2-5
Benchmarks (7)
Reasoning 3
Cite this model
BibTeX · APA
BibTeX
@misc{frontier-kimi-k2-5,
title = {Kimi K2.5},
author = {{Moonshot AI}},
year = {2026},
note = {Frontier Benchmarks AI atlas. Accessed 2026-05-08},
url = {https://frontierbenchmarks.com/models/kimi-k2-5}
} APA
Moonshot AI (2026). Kimi K2.5 [Large language model]. Frontier Benchmarks AI. Retrieved 2026-05-08, from https://frontierbenchmarks.com/models/kimi-k2-5
Citation reflects the atlas page, not the original model paper. For the paper, see the "Resources" section above.