K2 Think
The UAE's open-source reasoning system from MBZUAI and G42, running at up to 2,000 tokens per second on Cerebras hardware.
Overview
K2 Think is an open-source reasoning AI system developed by MBZUAI's Institute of Foundation Models with G42 in Abu Dhabi. The 32B model rivals reasoning models many times its size on math benchmarks while running at up to 2,000 tokens per second on Cerebras inference. K2 Think V2, a 70B system released in early 2026, is fully open end-to-end, from pre-training data through reasoning alignment, with weights, code, and training data published.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 12 Sep 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with K2 Think
- Run a 32B open reasoning model that leads open-source maths results — around 90.8% on AIME 2024
- Get up to ~2,000 tokens per second on Cerebras hardware
- Audit K2 Think V2 (70B) fully: weights, code and training data are published
- Build sovereign reasoning capability outside the US and Chinese model duopoly
- Read the security reporting before deploying: it was jailbroken within hours of launch, and that dominated the coverage
- Add your own guardrails rather than relying on the model's alignment, for the same reason
What this is for: An open-source reasoning system optimized for math and logic; the compact 32B model rivals far larger models, and K2 Think V2 (70B) is open end-to-end.
Who this is for: Researchers and developers who want a fast, fully transparent reasoning model to inspect and self-host.
Availability: Free and open-source; weights, code, and training data published on Hugging Face. A UAE sovereign model from MBZUAI and G42.
What people say
K2 Think got a strong technical debut and an immediate cautionary tale. On paper the 32B model punches far above its size, leading open-source math results (reported around 90.8% on AIME 2024, 73.75% on HMMT 2025, 71.08% on GPQA-Diamond) and running at up to ~2,000 tokens/second on Cerebras hardware, with MBZUAI and G42 pitching it as comparable to much larger reasoning models. Researchers who value auditability have welcomed the full openness of K2 Think V2 (70B), which publishes its weights, code, and training data.
But the story that dominated the security press was the jailbreak, which happened within hours of release. Adversa AI showed that the model's headline feature, transparent step-by-step reasoning, could be turned against it via partial prompt leaking: each refused attempt exposed defensive rules in the visible reasoning, which an attacker then explicitly countered, walking the model into producing malware and other illicit instructions. Dark Reading and others documented a slate of weaknesses: over-refusal, the jailbreak, and prompt extraction. So the honest read is a fast, transparent, benchmark-strong sovereign model whose transparency is simultaneously its selling point and a demonstrated safety liability; treat the vendor's benchmark leadership claims as awaiting broader independent replication.
Summary of public user & expert reviews, compiled by RECATOOLS.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including K2 Think's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with K2 Think unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to K2 Think directly →
Spotted something out of date? Suggest an update →
Alternatives to K2 Think
More in LLMs & Chat
