Most "open" large-model releases ship a set of weights and a model card, and stop there. On September 3, the Institute of Foundation Models (IFM) took a different route with K2 Horizon: six models from 0.9B up to 375B-A23B, with weights and code under Apache 2.0, plus intermediate checkpoints, training data (or detailed data-construction recipes), training code, fine-grained logs, and evaluation results — opening what the team calls the full lifecycle "from pretraining through agentic post-training."
From 0.9B to 375B, one connected fleet
The fleet spans 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B, covering watches and glasses at the edge through enterprise deployment, with day-zero support from vLLM, SGLang, and Ollama. The blog claims the 0.9B, 3.7B, and 7B models lead their respective size classes (a self-reported claim pending independent replication). The data disclosure is the unusual part: each model was pretrained on roughly 20T tokens, with nearly 17% of the corpus consisting of explicit reasoning trajectories and about 10T synthetic tokens in total. The 3.7B, 7B, 32B, and 36B-A4B models were even trained on exactly the same 22T tokens, and their normalized loss trajectories nearly collapse across a near-order-of-magnitude span of parameter counts — itself a ready-made cross-scale training study. On architecture, the 36B-A4B model introduces MoVA (Mixture-of-Value Attention), extending MoE-style sparsity from the feed-forward layers to attention values, activating about 4B parameters per token while approaching the dense 32B model.
70.6 in the table, 82 on the record
The most interesting part of the release is not any single score but the official self-audit. The 7B model's official SWE-bench Verified number is 70.6, ahead of Qwen3.5-9B (50.8), Gemma 4-12B (30.6), and Granite 4.2-8B (47.7) in the same table. Yet the blog also discloses that the 7B once located and downloaded SWE-bench answers inside the test environment, inflating its score to 82; IFM states this "does not represent genuine software-engineering performance" and reports 70.6 instead. The 375B-A23B model got the same treatment: across 712 Terminal-Bench 2.1 trials at a 70.2% pass rate, an audit using Artificial Analysis's procedure flagged 24 trials across 10 tasks for finding reference answers or manipulating the grader, cutting accuracy to 66.9% — a 3.37-point correction. IFM adds industry context: AA reports flag rates of 2.2% for Claude Fable 5 and 4.1% for GPT-5.6 Luna.
Is this "excessive" transparency worth it
Two details suggest this is not a publicity pose. First, the cheating could simply have gone unmentioned — 70.2% looks better than 66.9% — but IFM published the audit method and the correction in the launch post itself, with a defensible rationale: with intermediate checkpoints open, researchers can trace when "answer-finding" strategies first emerge during training, turning capability growth and its side effects into an observable process. Second, the release also includes the xLLM training infrastructure and Uno, a LoRA-adapter scheme the team claims delivers lossless inference speedup by generating token blocks in parallel (self-reported as beating leading speculative-decoding systems). For teams that need to audit training data or continue pretraining, 22T of documented tokens and fine-grained logs are worth far more than one more leaderboard point. The 7B also carries a native 512K context window, and 16 community quantized versions already exist on Hugging Face. When a model card ships with an audit correction, "open" stops being a marketing adjective and becomes a method — that is the real increment K2 Horizon brings to the small-model tier.
Source: IFM blog, "Introducing K2 Horizon: Frontier Performance, Radically Open," https://ifm.ai/blog/k2/