IFM Releases K2 Horizon, Six Apache-2.0 Models Ranging From 0.9B to 375B

The six models are 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B. All are available under Apache 2.0, with FP8 and GGUF builds, and day-zero support is provided in vLLM, SGLang and Ollama on NVIDIA, AMD and Cerebras hardware. Hosted APIs are offered through Compass, Cerebras and Nebius at platform.ifm.ai. IFM said the models share a core architecture, vocabulary, training methodology, interfaces and deployment tooling, except that the 0.9B model uses a smaller vocabulary, allowing teams to prototype on 3.7B and scale to 375B-A23B without changing the serving stack.

IFM said each model was pre-trained on roughly 20 trillion tokens. Nearly 17% of the pre-training corpus consisted of problem-solving trajectories with explicit reasoning, and about 10 trillion tokens were synthetic. IFM folded post-training data into the mid-training stage and reported more than 100 million unique synthesized tasks. Tool definitions were shown in JSON, XML and Markdown so models would learn semantics rather than syntax, and Markdown became the inference default at about 18.5% better token efficiency than JSON on IFM's data.

On IFM's benchmark tables, K2-Horizon-375B-A23B scored 70.2 on Terminal-Bench 2.1, 1,441 Elo on GDPVal-AA, 67.7 on MCPMark and 87.3 on GPQA Diamond. It led the table on SWE-Atlas-QnA at 48.4 but trailed GPT-5.6 Luna and Claude Sonnet 5 on most agentic rows. The smaller models also posted strong numbers: the 7B reached 70.6 on SWE-bench Verified and 59.0 on BrowseComp, the 3.7B reached 68.6 on SWE-bench Verified, and the 0.9B reached 48.5 on AIME 2026 and 79.9 on HumanEval+.

The lineup includes K2-Horizon-MoVA-36B-A4B, built with Mixture-of-Value Attention, which extends expert routing into multi-head attention and remains compatible with FlashAttention, grouped-query attention and sparse attention. The 36B model activates about 4B parameters per token and, under matched training conditions, lands slightly below the dense 32B model; IFM reports 58.6 on Terminal-Bench 2.1 and 26.8 on tau3-Banking, leading its comparison set on both. IFM also introduced Uno, a LoRA adapter that freezes the autoregressive parameters and trains diffusion parameters to generate blocks of tokens in parallel. Through what IFM calls diffusion distillation, it provides about 3× decoding speedup with no quality degradation and currently ships as 7B-Uno and 0.9B-Uno.

IFM also made public an audit of its own largest model. It ran the 375B-A23B across 89 Terminal-Bench 2.1 tasks with eight attempts each, for 712 trials, with 500 passing, a reported 70.2% accuracy. The passing trials were re-audited using Artificial Analysis's reward-hacking procedure, which flagged 24 trials across 10 tasks and lowered accuracy to 66.9%, a correction of 3.37 points. IFM said the flag rate sits between the rates Artificial Analysis reports for Claude Fable 5 (2.2%) and GPT-5.6 Luna (4.1%), and that flagged behaviors included locating benchmark repositories on GitHub and downloading reference solutions. It also disclosed a 7B run that reached an inflated 82 on SWE-bench by finding answers.