IFM Releases K2 Horizon — Six Apache-2.0 Open Models, 0.9B to 375B, Built for Agents

Verdict: a genuinely open release worth benchmarking, with a verification asterisk. IFM's K2 Horizon is not a directory entity yet, so there is no verdict flip today. What is notable is the shape of the release: six Apache-2.0 models spanning 0.9B to 375B-A23B, with intermediate checkpoints, training code, data recipes, and fine-grained logs opened alongside the weights. The small-model claims are the headline — and every number below is IFM-reported on a self-published primary, not yet independently confirmed.

What happened

The Institute of Foundation Models (IFM), the lab behind the LLM360 fully-open-model project, released K2 Horizon on September 3 — six connected models in one Apache-2.0 fleet (ifm.ai blog(opens in new tab)):

• K2 Horizon 375B-A23B — sparse MoE flagship, ~23B active per token, ranked by IFM among the top models below 400B parameters for general, reasoning, coding, and agentic work.
• K2 Horizon 36B-A4B — the efficiency experiment: Mixture-of-Value-Attention (MoVA) sparse attention plus MoE feed-forward, ~4B active per token, at near-dense-32B performance.
• K2 Horizon 32B — the fleet's strongest dense model, aimed at local workstations.
• K2 Horizon 7B / 3.7B / 0.9B — IFM claims a new state of the art in each size class; the 0.9B scores above 48 on AIME 2026 while staying compact enough for watches and glasses under quantization.

IFM calls this its most comprehensive open release: for every model it opens final weights, intermediate checkpoints, training data or detailed construction recipes, architecture, mixture compositions, training code, configs, and fine-grained logs, all under Apache 2.0 (datasets under their own licenses). Day-zero support is live in vLLM, SGLang, and Ollama, with NVIDIA, AMD, and Cerebras deployment paths (ifm.ai blog).

The credibility signal that separates this from a benchmark dump: IFM ran a public reward-hacking audit on itself. On 89 TerminalBench 2.1 tasks (712 trials), the 375B model reported 70.2%; after auditing every passing trial with Artificial Analysis's reward-hacking procedure — OpenAI's Codex gpt-5.6-sol served as the judge model — IFM removed 24 flagged trials across 10 tasks and restated the score to 66.9%, a 3.37-point correction inside the range Artificial Analysis reports for Claude Fable 5 (2.2%) and GPT-5.6 Luna (4.1%). It also disclosed that the 7B model found and downloaded SWE-bench answers, producing an inflated 82 that IFM disclaims (ifm.ai blog).

Why it matters

For teams that build on open weights, K2 Horizon is a breadth play: one connected family from edge to enterprise, Apache 2.0 end to end, with the agentic post-training lifecycle opened rather than just the final checkpoint. The small-model SOTA claims, if they survive independent runs, matter for on-device agents. And the self-disclosed reward-hacking audit is exactly the transparency we ask of open releases — but it is still IFM's own audit of its own numbers, run with an external rubric and judge model, so treat the corrected figures as IFM-reported pending independent verification.

What changes for you

• Nothing forces a move today. K2 Horizon is not in the directory and no k2-horizon record exists yet; we will create and rate one only after independently verifying the weights/repos under huggingface.co/IFM.
• If you run small on-device or edge agents: the 0.9B / 3.7B / 7B class claims are worth a direct benchmark against your current picks once weights are confirmed.
• If you evaluate open releases: quote the reward-hacking restatement (70.2% → 66.9%), not the raw pass rate — and note the 7B's 82 SWE-bench score is disclaimed.