IFM Ships Six Open LLMs and the Entire Lab Notebook Behind Them

What happened

IFM released K2 Horizon on September 3, 2026: a fleet of six open-weight LLMs from 0.9B to 375B parameters, all Apache-2.0 licensed, with the full training pipeline — checkpoints, data recipes, code, configs, and logs — published alongside. It is the first open model family to expose its entire development process, including agentic post-training, as a public artifact.

Context

IFM introduced the fully open principle in its 2023 LLM360 paper and has shipped open models every year since. The gap this release targets is not model quality — open weights have been available for two years — but process transparency: what training data went in, how post-training was staged, and what the intermediate checkpoints look like. Prior open releases typically shipped final weights and a model card. K2 Horizon ships the whole lab notebook.

How it works

Six sizes, two architectures. The 375B-A23B is a sparse Mixture-of-Experts model: 375 billion total parameters, roughly 23 billion active per token. The 36B-A4B introduces a new Mixture-of-Value-Attention (MoVA) mechanism and outperforms some larger models on capability per active parameter. The 32B is the fleet's most powerful dense model. Below that, 7B, 3.7B, and 0.9B are dense models targeted at phones and constrained hardware such as watches and glasses; the 0.9B uses a reduced vocabulary. All six share core architecture, tokenizer (except the 0.9B), training pipeline, interfaces, evaluation infrastructure, and deployment tooling. Every model ships with quantization support. Published evaluations cover AIME 2026 (the 0.9B scores above 48), SWE-bench, BrowseComp, and TerminalBench, though IFM notes TerminalBench remains difficult for the smallest models.

Our read

The headline number — 0.9B clearing 48 on AIME 2026 — gets the clicks, but the structural story is the shared backbone. Six models that share architecture, vocabulary, interfaces, and tooling means a routing layer is trivial: you hot-swap between a 7B for fast task parsing and a 32B or 36B-A4B for heavier reasoning without touching tokenization or the surrounding code. For a studio running ComfyUI, that is a model-selection knob, not a rewrite.

What the press release does not say: these are text-generation LLMs. They complement the image and video diffusion checkpoints a studio already runs; they do not replace them. The agentic post-training data and recipes IFM is publishing are where the real leverage sits for a small team — fine-tune a 7B to call ComfyUI nodes, manage asset APIs, structure shot lists, and you have a pipeline orchestrator that can plausibly run on a single high-end consumer GPU. But the sources give no inference-speed numbers, no context-window sizes, and no VRAM floor per quantization level, so that "runs on my one GPU" claim is plausible, not confirmed.

The second-order effect is more interesting. If every open family starts publishing training data recipes and intermediate checkpoints as a matter of course, the differentiation shifts from which model to which data mix. The model card becomes a starting point, not the product.

What this changes

For a studio that already runs local models: pull the 7B for prompt parsing and scene-graph structuring in a ComfyUI workflow; keep the 32B or 36B-A4B at reduced quantization for shot-list generation and VFX-logic reasoning where the 7B falls short. The shared tokenizer means one interface layer across the swap. The agentic post-training recipes are the piece to watch: if they are genuinely reusable, a 7B fine-tuned on ComfyUI node-calling becomes a domain-specific orchestrator without reaching for a 375B model. Nothing changes for the diffusion side of the stack. Nothing changes yet for the 0.9B unless you are shipping into watches or glasses.

License

Models and code are Apache-2.0. Commercial use, modification, and redistribution are permitted without royalty or attribution beyond the standard Apache notice. Training datasets carry their own licenses, including ODC-BY, so check the specific dataset terms before redistributing any derived data.

Key takeaways

• Six sizes, shared architecture and tooling — a routing layer between them is a config change, not a rewrite.
• 0.9B scores above 48 on AIME 2026; the smallest models still struggle on TerminalBench.
• 36B-A4B's MoVA mechanism and 375B-A23B's sparse MoE target different deployment tiers, from local workstation to enterprise.
• Full training-pipeline transparency (checkpoints, data recipes, code, logs) is new for open-weight releases and shifts the competitive axis toward data composition.
• These are text LLMs; they orchestrate and reason, they do not generate images or video.

Sources

1. K2 Horizon: A connected fleet of six open models — tier 3