K2 Horizon 7B

IFM K2 Horizon 7B (Apache 2.0), released 2026-09-03 — a dense 7B-core model (9B parameters stored, 512K context) for local workstations, with fully open training data and code. Runs today on vLLM, SGLang and IFM's llama.cpp fork; not yet on stock Ollama or LM Studio.

Type: local

Context: 524,288 tokens

Use cases: general, coding, reasoning

Run K2 Horizon 7B locally

There is no official Ollama tag for K2 Horizon 7B yet. The model uses a new k2-horizon architecture that stock Ollama, LM Studio and upstream llama.cpp cannot load until IFM's llama.cpp pull request lands — "ollama run k2-horizon" will fail. Use IFM's llama.cpp fork or vLLM instead. The official GGUF is BF16 (about 18 GB); community 4-bit quants built for the fork are 5–6 GB and fit an 8 GB GPU at short context.

Min VRAM: 8 GB

llama.cpp (MBZUAI-IFM fork, OpenAI API on :8080)

git clone -b model/K2Horizon https://github.com/MBZUAI-IFM/llama.cpp
cd llama.cpp && cmake -B build && cmake --build build -j
./build/bin/llama-server -hf IFM/K2-Horizon-7B-GGUF -ngl 99 -c 32768

vLLM (GPU with ~24 GB, OpenAI API on :8000)

vllm serve IFM/K2-Horizon-7B \
  --trust-remote-code \
  --dtype bfloat16 \
  --max-model-len 131072 \
  --reasoning-parser k2_horizon \
  --enable-auto-tool-choice \
  --tool-call-parser k2_horizon

Limits checked against official docs on 2026-10-10. Rates change without notice. Not affiliated with providers unless noted. Verify before production use.