← Back to directory
Moonshot AI
Kimi K2 Thinking
Released Nov 6, 2025 · knowledge cutoff 2025-04
- Status
- Superseded
- Location
- China
- Type
- LLM
- Modality
- Text
- Context window
- 262K
- Max output
- 66K
- Speed
- —
- Price ($/MTok in / out)
- $0.60 / $2.50
- Cost per task
- —
- Open weights
- Yes
Benchmarks6 of 10 reported
MMLU-Proretired
84.6%
GPQA Diamond
84.5%
SWE-bench Verified
71.3%
AIME (latest)retired
94.5%
Humanity's Last Exam
23.9%
LMArena Elo
1417
SWE-bench Pro
—
Terminal-Bench 2.1
—
GDPval-AA v2
—
ARC-AGI-2
—
Strengths & weaknesses
Strengths
- Chains hundreds of tool calls without losing the goal — rare stamina for an open model
- Research-style depth close to closed models of its moment
- Weights available, so the entire agent loop can run in-house
Weaknesses
- Slow, and it deliberates on nearly everything whether or not it needs to
- Needs low-precision serving to be practical, which is fiddly to set up correctly
The strongest open reasoning model as of late 2025. Licence is Moonshot's Modified MIT: MIT plus a 'Kimi K2' UI-attribution clause above 100M MAU or US$20M monthly revenue, classified permissive.
For developers
API model strings
- moonshot
kimi-k2-thinking
Licence
Modified MIT License · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor