GitHub radar

K2-Horizon-7B: fit a whole project in one coding prompt

IFM (Institute of Foundation Models) released a 9B model with a native 512K-token context and 70.6% on SWE-bench Verified — runs via Ollama on a regular laptop.

01IFM/K2-Horizon-7B 1827.4k downloads/mo9.0B paramstext-generation

K2-Horizon-7B is a language model from IFM (Institute of Foundation Models) with a native 512K-token context window — enough to fit an entire mid-sized codebase or a large repository in one prompt. It was designed for agentic use: tool calling, terminal tasks, browser use, and multi-step code work. On SWE-bench Verified, the standard benchmark for patching real-world code, it scores 70.6% — well above other models in the same weight class, which score 30–50%. Training data, recipe, and evaluation code are fully public. Quantized versions run via Ollama and LM Studio.

Why a vibe-coder should care

The 512K context means you can hand the model an entire project at once — no chunking, no losing the thread. For simple coding tasks without a subscription, it's a solid local option. For real agentic work, frontier models still hold the edge — but the 70.6% SWE-bench score makes this worth trying.

How to install

Copy this and send it to your agent — Claude Code, Codex, any of them:

Set up K2-Horizon-7B via Ollama: https://huggingface.co/IFM/K2-Horizon-7B — find the right quantized variant for my machine and show me how to give it a coding task.

A regular laptop with 16 GB of RAM or more — the 4-bit version takes around 5 GB.

Open on Hugging Face