Runs on ≤32 GB · Hugging Face ↗
Not independently scored by Artificial Analysis
Released September 2026 by MBZUAI's Institute of Foundation Models, Apache-2.0. A mixture-of-experts that stores 37.4B parameters and reads about 4B per token: 48 layers, 100 experts with 8 active plus 1 shared. The new part is the attention. "MoVA" routes the hidden state through a pool of 64 value matrices and uses 4 per token, so the values are a mixture of experts too, not just the feed-forward block. That is also why nothing off the shelf runs it: the architecture id in the GGUF header is k2-horizon, and no released llama.cpp knows that name. Note this is a family, not one model - 0.9B, 3.7B, 7B and 32B dense, this 36B-A4B, and a 375B-A23B. The 375B is the one Artificial Analysis scored; it is 758 GB and fits nothing we have, so no number on this page comes from it. ACTUALIZADO el 2026-10-03: ya SI lo carga llama.cpp, y lo hemos medido con el protocolo oficial, tres tandas sobre el GGUF oficial de IFM. La frase de que ningun runtime lo cargaba vale solo hasta septiembre de 2026.
BigCodeBench-Hard pass@1 27% (40/148), via official protocol · llama.cpp Q4_K_M · one request at a time · a shared cluster GPU · mean of 3 runs, range 0.7 pp. The brutal counterpart to HumanEval — where HumanEval saturates near the top, BCB-Hard spreads the field, so this is the number that actually separates coding ability. 3 of 148 answers hit the token limit (effective ceiling 98%)
Our own runs on this model. Each number carries how many problems it is over and how many times we ran it — these are not comparable with each other, and none of them is on the same scale as the third-party indices on the leaderboard.
| test | score | over | conditions |
|---|---|---|---|
| BigCodeBench-Hard | 27.3% | n=148 · 3 run(s) | official protocol · llama.cpp Q4_K_M · one request at a time · a shared cluster GPU · mean of 3 runs, range 0. |
Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈6 GB at 32K, fp16). Read from the GGUF header, but this architecture has never been checked against llama.cpp — treat it as an estimate.
| Quantization | Size |
|---|---|
| Q4_K_M | 22.4 GB |
| Q5_0 | 25.9 GB |
| Q5_K_M | 26.4 GB |
| Q6_K | 30.8 GB |
| Q8_0 | 39.8 GB |
| BF16 | 74.9 GB |
Do not start with a GGUF. We tested it: llama-cli build 8950 on the smallest quant
answers "error loading model architecture: unknown model architecture: 'k2-horizon'".
Ollama and LM Studio use the same loader, so they fail the same way. The 31
quantizations on the Hub are all unusable for now; the llama.cpp work sits in a draft
branch of IFM's own fork, not upstream.
The route that works today is Hugging Face transformers from safetensors, with
trust_remote_code=True - the repo ships modeling_k2_horizon.py and an auto_map. We
loaded it on transformers 5.16.1 with torch 2.13; flash_attention_2 was not available
for our build, so it fell back to sdpa, which works. Budget about 75 GB for bf16
weights. The model card also lists vLLM and SGLang.
27.7% over three runs (27.7 / 27.0 / 27.0), the best of its family by six points over the 7B. With 4B active it runs fast for its size. Sobre el GGUF: esta ficha apunta al oficial de IFM, pero el espejo de terceros NANI-Nithin sirve EL MISMO fichero, byte a byte: 22.368.011.616 bytes y sha256 513dd785...4cbd en los dos. Comprobado el 2026-10-03 contra el Hub, y es el fichero que se midio. No hace falta cambiar el repo.