hako-sho

Mini tier (3B) of the hako local model family from Mithraeum. A stock wrap of Qwen2.5-Coder-3B-Instruct, converted to the MLF2 container and run by the from-scratch hako engine (hakm) — no llama.cpp, no ggml, no ollama at runtime.

This is the default local model for the hako-code agent and hako-edit editor.

What this is

• Base: stock Qwen2.5-Coder-3B-Instruct — no weight fine-tune yet. A real fine-tune on Mithraeum data will earn a versioned id (hako-sho-v0.0.1).
• Format: MLF2 (the hako engine's container; k-quant blocks copied verbatim from the source GGUF). Loaded via mmap, int8-activation × int4-weight fast path.
• Template: ChatML (<|im_start|> / <|im_end|>).

Use it

# via the hako-code agent (auto-default; pulls this repo):
hako                       # then  :pull hako-sho

# or standalone with the engine CLI:
hakm hako-sho.mlf2 --sys "You are hako." "explain ring buffers in C"

The agent resolves :pull hako-sho to https://huggingface.co/mithraeum/hako-sho/resolve/main/hako-sho.mlf2.

License

⚠️ Qwen RESEARCH license — non-commercial use only (the 3B base is research-licensed, not Apache). See LICENSE and NOTICE. The hako engine code is GPL-3.0; these weights keep their required upstream license.