K2 Horizon 7B
The mid-sized reasoning model of the K2 Horizon family, released in September 2026 by the Institute of Foundation Models (IFM, MBZUAI) under the Apache 2.0 license. It is open far beyond the weights: the training data and recipe are public, as are most intermediate checkpoints and the training logs; the code and technical report are announced for the end of September. It reads 524,288 tokens of context. The name says 7B, but Hugging Face counts 9 billion parameters in the published weights (IFM describes a 7B core). Artificial Analysis scores it 21 on its Intelligence Index, where the median for open models of similar size sits at 8. IFM reports 70.6 on SWE-bench Verified, against 50.8 for Qwen3.5-9B in its own table.
Strengths
- Scores 21 on the Artificial Analysis Intelligence Index against a median of 8 for open models its size: an independent measurement, not a vendor figure
- Open all the way down: public training data and recipe, intermediate checkpoints from pretraining through SFT, training logs, all under Apache 2.0
- 524,288 tokens of native context, enough to load a large set of documents without chunking
- Served directly by vLLM and SGLang with recipes validated by IFM, tool calls included
Limitations
- Verbose: 160 million tokens generated to get through the Artificial Analysis index, against 76 million for the median of its size. Locally, you pay for it in waiting time
- "7B" in the name, 9 billion parameters in the weights: the only official GGUF is BF16 and weighs 18 GB, out of reach for a 16 GB machine
- llama.cpp does not know the architecture yet (PR in progress): meanwhile the GGUF repository points to the IFM fork, which you build yourself
- The card declares English only, and neither the code nor the technical report is published yet: keep an eye on how well it really handles other languages
Best for
- Developers who want an open, local reasoning model, with a GPU or at least 18 GB of free memory
- Teams that have to justify where a model comes from: data, recipe and checkpoints can be inspected
- Curious builders who want to understand how a model acquires its abilities, stage by stage, through the intermediate checkpoints