MBZUAI's IFM ships K2 Horizon, six Apache 2.0 models up to 375B

IFM released K2 Horizon, six Apache 2.0 models from 0.9B to a 375B MoE with 512K context, and has already posted intermediate checkpoints and five training datasets; its code repository is due by end of September, and hosted access runs through Cerebras, AWS and Nebius.

Key points

• Repos: IFM/K2-Horizon-0.9B, -3.7B, -7B and -32B (dense), plus IFM/K2-Horizon-MoVA-36B-A4B and IFM/K2-Horizon-375B-A23B (MoE); the 7B and 0.9B also get 'Uno' adapters, adapter-only LoRA add-ons that give the base model a diffusion decoding path for faster generation (see the 7B-Uno card and the paper it links), and GGUF repos exist for the 0.9B, 3.7B, 7B, 32B and 36B-A4B.
• Flagship: 375B total, 23B active per token, native 524,288-token (512K) context. The model card's vLLM recipe caps --max-model-len at 131,072; its SGLang recipe, which the card says is validated on 8x H200, sets no context flag.
• Apache 2.0 across the model family. Besides the final checkpoint, the 375B repo holds intermediate checkpoints from every pretraining, midtraining and SFT stage plus the merged RL checkpoint; the card's release table puts the five RL expert checkpoints at mid-September and the code repository and technical report at end of September. Five K2 Horizon training datasets are already public on Hugging Face (listed on the TxT360-v2 card). The press release and IFM's Hector Liu (in The National) describe each model as shipping with its data, code and methodology.
• IFM's own reported 375B scores, from the card's benchmark table: Terminal-Bench 2.1 70.2, SWE Bench Pro (strict) 42.6, Toolathlon Verified 65.3, SciCode 42.7, GDPVal-AA 1,441 Elo. The same table puts GLM 5.2 at 77.9 and GPT 5.6 Luna at 80.9 on Terminal-Bench 2.1.
• Smaller models: the 36B-A4B, which uses what IFM calls Mixture-of-Values attention, reports 58.6 on Terminal-Bench 2.1 and 80.8 on GPQA Diamond; the 3.7B dense reports 68.6 on SWE-bench Verified and 25.1 on Terminal-Bench 2.1. Both carry 512K native context; the 0.9B is the exception at 128K.
• Availability: Hugging Face now, with vLLM and SGLang support, per the press release. IFM's GGUF builds keep the full BF16 weights and need IFM's own llama.cpp fork until upstream support lands; The National also lists Ollama and Unsloth. Hosted API through Compass, Cerebras, AWS and Nebius. No pricing published on any page fetched.

“Open source is much more than open weights.” — Eric Xing, IFM founder and MBZUAI president, in the launch press release

Why it matters

For a studio that self-hosts, the 36B-A4B and the 32B are the realistic targets (the 32B card says its final checkpoint is still to come): the 36B-A4B runs only 4B active parameters per token, which keeps inference cheap, and GGUF builds are up, though for now they need IFM's llama.cpp fork. The 375B is a credible open agentic model, but in IFM's own Terminal-Bench 2.1 comparison it trails GLM 5.2 and the closed frontier, so treat it as a fallback rather than a drop-in replacement. Checkpoints and five training datasets are already public; watch whether the code repository follows by IFM's end-of-September date, since code is part of what separates a fully open release from plain open weights.

This is an AI-written summary of the reporting credited above and the other sources linked in the text, read and edited by Nicholas before publishing. The facts belong to those sources and the quote to Eric Xing, via IFM's press release; the wording is ours. Read the original.