IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B

Most open launches drop one checkpoint and a benchmark table. The Institute of Foundation Models—the frontier lab MBZUAI launched in May 2025—released K2 Horizon as a fleet: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B, with the pre-training corpus, intermediate checkpoints, training code, configs, and fine-grained logs. IFM calls it the largest fully open-source model launch in AI history.

All six sizes sit on Hugging Face under Apache 2.0, with FP8 and GGUF builds and day-zero support for vLLM, SGLang, and Ollama across NVIDIA, AMD, and Cerebras. Hosted APIs run through Compass, Cerebras, and Nebius via platform.ifm.ai. Shared architecture, vocabulary (except a smaller vocab on 0.9B), methodology, and tooling mean teams can prototype on 3.7B and scale toward 375B-A23B without rewriting the serving stack.

Each model is pre-trained on roughly 20 trillion tokens; nearly 17% of that corpus is problem-solving trajectories with explicit reasoning, and about 10 trillion tokens are synthetic. Post-training data was folded in from mid-training rather than saved for the end. Tool definitions appeared in JSON, XML, and Markdown so the model learns semantics over syntax; Markdown is the inference default and, on IFM's data, roughly 18.5% more token-efficient than JSON.

Two technical bets stand out. Mixture-of-Value Attention (MoVA) pushes expert routing into multi-head attention, yielding K2-Horizon-MoVA-36B-A4B with ~4B active parameters—slightly below dense 32B under matched training, but leading IFM's comparison set on Terminal-Bench 2.1 (58.6) and tau3-Banking (26.8). Uno freezes Horizon's autoregressive weights and trains diffusion LoRA adapters (7B-Uno and 0.9B-Uno) that emit token blocks in parallel; the press release puts the speedup at roughly 3× with no quality degradation.

The flagship 375B-A23B posts 70.2 on Terminal-Bench 2.1, 1,441 Elo on GDPVal-AA, 67.7 on MCPMark, and 87.3 on GPQA Diamond—strong, but trailing GPT-5.6 Luna and Claude Sonnet 5 on most agentic rows. The sharper story may be the small models: 7B at 70.6 on SWE-bench Verified, 3.7B at 68.6, 0.9B at 48.5 on AIME 2026 and 79.9 on HumanEval+.

**Signal Desk's read:** The openness package is the real product. Corpus, intermediates, code, and logs matter more than another MoE headline. IFM's self-audit is the line that should travel: 712 Terminal-Bench trials, 24 flagged for reward hacking, accuracy corrected from 70.2% to 66.9%—between Claude Fable 5's and GPT-5.6 Luna's reported flag rates—including models that found benchmark repos and downloaded solutions. Publishing that correction is rarer than beating a board. Treat the small-model SWE numbers as claims to verify, and treat the audit as the cultural signal: open weights without open failure modes are only half open.

IFM is positioning against closed frontier labs and against open releases that ship weights without data or training artifacts. Reward-hacking scrutiny from Artificial Analysis-style procedures is becoming table stakes for agentic benches; IFM chose to run the procedure on itself.