MBZUAI's K2 Horizon 375B Beats Rivals on Real-World Agent Tasks

Abu Dhabi's K2 Horizon 375B A23B is the flagship of a six-model open-weights fleet from MBZUAI's Institute of Foundation Models, scoring 47 on the Artificial Analysis Intelligence Index — a 30-point jump over its 70B dense predecessor, K2 Think V2 — while targeting agentic workloads rather than trivia recall.

A fleet, not just a flagship

K2 Horizon spans multiple size classes across reasoning, mathematics, coding, and agentic tasks. The 0.9B model targets highly constrained environments like watches and glasses; 3.7B and 7B bring advanced capabilities to phones and on-device apps; a dense 32B and sparse 36B-A4B cover local hosting and on-premise servers; and the 375B-A23B tops the fleet for demanding enterprise deployments. The API is live today through IFM's inference partners including Compass, Cerebras, AWS, and Nebius. Models and code are released under Apache 2.0.

Under the hood

The flagship replaces the previous dense transformer with a sparse mixture-of-experts design. Key specs:

• 375B total parameters, 23B active per token (MoE routing)
• 524K token context window (roughly 786 A4 pages of 12pt text)
• Text in, text out (no vision)
• Reasoning model with extended chain-of-thought
• Apache 2.0 license

That puts it in direct architectural company with MiniMax-M3, which also runs 23B active parameters over a larger sparse pool.

Where it wins, where it loses

Against MiniMax-M3 — the closest Intelligence Index neighbor at 45 — K2 Horizon skews toward practical work rather than exam-style reasoning. It leads on GDPval-AA, a real-world knowledge work benchmark, with an Elo of 1430 versus 1380, and dominates τ³-Banking (34.2% versus 15.3%), a tool-use benchmark simulating a banking agent. It trails on GPQA Diamond (87.3% versus 92.9%) and Humanity's Last Exam (32.0% versus 39.0%), so the hardest graduate-level science and reasoning questions still favor competitors.

The hallucination story

On AA-Omniscience, a knowledge reliability benchmark, K2 Horizon attempts only 40% of questions and declines the rest. That produces a 26% hallucination rate — among the lowest measured — while raw accuracy sits at 18%, essentially flat versus K2 Think V2. The composite Omniscience Index climbed from -40 to -3 because the model stops confidently guessing wrong rather than reaching for a plausible-sounding answer.

For retrieval pipelines or agents that need to hand off to a human when uncertain, that abstention behavior is more useful than a few extra points of raw recall.

Open by design

The K2 line has built its reputation on transparency, not just weight availability. K2 Think V2 shipped with weights, training code, data composition, mid-training checkpoints, and evaluation frameworks, positioning it as one of the few fully open-source models that is both performant and reproducible — distinct from the semi-open tier of Chinese open-weight releases. Artificial Analysis ranks K2 Think V2 among the most open models on its Openness Index; K2 Horizon is expected to be added once supporting documentation is finalized.

Lineage and context

K2 Think launched in September 2025 with 32 billion parameters for reasoning-intensive workloads. By January 2026, the team had scaled to K2 Think V2 at 70 billion parameters. Horizon is the first move to sparse MoE, doubles the context from 262K to 524K tokens, and reframes the family around agentic capability rather than pure reasoning benchmarks.

Who should use it

K2 Horizon 375B A23B is worth deploying if you are:

• Building tool-using agents where τ³-Banking-style multi-turn interactions matter
• Running knowledge-work automation — drafting analyses, presentations — where GDPval-AA is representative
• Deploying in environments where hallucination cost is high and abstention is acceptable
• Needing an Apache 2.0 model with fully open training artifacts for compliance or fine-tuning
• Working with long context (up to 524K tokens) for document-heavy RAG

If your workload centers on frontier scientific reasoning or PhD-level knowledge benchmarks, MiniMax-M3, Qwen, or DeepSeek variants of similar size still hold an edge. For the agentic middle ground where most production LLM work lives, a 23B-active MoE that refuses to bluff is a genuinely useful addition to the open-weights menu.