IFM's K2 Horizon: six fully-open models, 0.9B–375B, Apache 2.0
MBZUAI ships weights, training data and code together—rare. The 7B tops Qwen on SWE-bench and runs day-one on Ollama and vLLM. Big-model data still lags.
Copy markdown
The fleet: six sizes, one license
MBZUAI's Institute of Foundation Models dropped six models—0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B—all under Apache 2.0, so you can self-host and ship commercially with no strings. IFM is billing it as the largest fully open AI release to date.
Weights, data AND code—not just a checkpoint
Unlike most 'open' drops, K2 also publishes the training data (the TxT360-v2 set), the pre/post-training code, and the recipes. That's the part that lets you actually reproduce the run or retrain on your own domain, not just run inference on a frozen model.
The 7B punches far above its weight
The 7B scores 70.6 on SWE-bench Verified (vs 50.8 for Qwen3.5-9B), 39.1 on Terminal-Bench 2.1 and 59.0 on BrowseComp—IFM claims state-of-the-art for its size class on coding and agentic tasks. A genuinely capable agent model that fits on a single GPU.
Run it tonight
Day-one support for vLLM, SGLang, Ollama and Unsloth, plus GGUF builds on the IFM Hugging Face org—so local inference and fine-tunes work out of the box. Hosted APIs are live via Cerebras, AWS and Nebius, though no K2 per-token pricing is published yet.
The flagship: 375B MoE, 524K context
The top model activates 23B of 375B params and takes a 524K-token window. Artificial Analysis rates it 47 on its Intelligence Index—well above the ~29 median for comparable open-weight models—and it leads GLM 5.2 on tool use (Toolathlon 65.3), though it trails on some other agent benchmarks.
The honest caveat: 'fully' is staggered
Only the 3.7B and 7B ship complete today. The 375B and 36B model cards say checkpoints, data and training code 'will be released,' and the 32B is a Stage-1 checkpoint with the final still pending. Open as a direction—but the big-model downloads are running ahead of the promise.