K2-Horizon 3.7B

About

Late in training, its lessons shifted toward agent and reasoning work while it learned to handle longer input. Training split it into several expert versions, which were then merged back into one model. It handles more than one tool-call format, and its released versions can be trained further.

How it compares

Of the 7 chat models Outmake recommends for an 8 GB Mac, it ranks 5th fastest by Outmake's estimate.

Outmake estimates Qwen3.5 4B writes about twice as fast, and Qwen3.5 4B also reads pictures, which this model can't.

On Hugging Face it was the least downloaded of those 7 models last month.

Limitations

If its reasoning gets cut off, the answer counts as a failure.

Good to know

An extra training stage used pairs drawn from its own looping answers to curb repetition.

Its training data, recipe, code and evaluation resources are all public.

Versions from partway through training are released, so people can study how its abilities grew.

Sources: IFM/K2-Horizon-3.7B model card, Outmake's speed estimates, LM Arena and Hugging Face. Every line comes from one of these.

Run It Anywhere

K2-Horizon 3.7B is an open model, so any app that runs GGUF files can run it. These take the Q4_K_M size from bartowski straight from Hugging Face.

• LM StudioMac, Windows, LinuxOpen
• JanMac, Windows, LinuxOpen
• MstyMac, Windows, LinuxOpen
• OllamaMac, Windows, Linuxollama run hf.co/bartowski/K2-Horizon-3.7B-GGUF:Q4_K_MCopy
• llama.cppMac, Windows, Linuxllama-server -hf bartowski/K2-Horizon-3.7B-GGUF:Q4_K_MCopy
• Docker Model RunnerMac, Windows, Linuxdocker model run hf.co/bartowski/K2-Horizon-3.7B-GGUF:Q4_K_MCopy