K2 Think
32B open reasoning model that rivals models 20x its size
About
K2 Think is a 32-billion-parameter reasoning model from MBZUAI's Institute of Foundation Models and G42, announced on September 9, 2025. It was built with long chain-of-thought fine-tuning, reinforcement learning with verifiable rewards, and agentic planning, and the team claims it matches flagship reasoning models up to 20 times its size. Weights, training data, code, deployment tooling and safety evaluations are all public, which is rare at this performance tier. It is reasoning-first rather than a general chat model, so it is less suited to casual conversation.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: leads all open-source models on AIME '24/'25, HMMT '25 and OMNI-Math-HARD; strong on LiveCodeBench v5 and GPQA-Diamond. 2,000 tokens/sec on Cerebras hardware.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Weights, training data, code, deployment tooling and safety evaluations all public; Strong launch coverage (Sept 2025) from MBZUAI/G42 announcements; Performance claims are team-reported; independent reproductions still limited. License: Described as fully open-source by MBZUAI; V2 is stated to be 100% sovereign and open from pre-training through post-training..
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Weights, training data, code, deployment tooling and safety evaluations all public
- · Strong launch coverage (Sept 2025) from MBZUAI/G42 announcements
- · Performance claims are team-reported; independent reproductions still limited