CrowQwen3.5 starts from Qwen/Qwen3.5-4B and is tuned toward local assistant work: coding notes, refactor plans, bug triage, shell-oriented workflows, and bilingual technical prompts.

Highlights

Korean + code + English + agent SFT mix.
Tool-call formatting support in the chat template.
Heretic-edited final export selected by internal refusal/KL probe.
9 MLX/oMLX oQ variants for Apple Silicon.

Pick A Directory

Goal	Recommended
Balanced default	`oQ4`
More fidelity	`oQ5`, `oQ6`, `oQ8`
Smaller local runs	`oQ2.7`, `oQ3`, `oQ3.5`
Extra compact choices	`oQ2`, `oQ2.5`

Start with oQ4 if you are unsure.

Quick Start

bash

huggingface-cli download \
  dawncr0w/crowqwen3.5-4b-agent-heretic-mlx-oq \
  --include "crowqwen3.5-4b-agent-heretic-oQ4/*" \
  --local-dir crowqwen-mlx-oq

python

from mlx_lm import load, generate

model, tokenizer = load("crowqwen-mlx-oq/crowqwen3.5-4b-agent-heretic-oQ4")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Draft a short CLI refactor plan."}],
    tokenize=False,
    add_generation_prompt=True,
)

print(generate(model, tokenizer, prompt=prompt, max_tokens=160))

Artifacts

Variant	Size	Smoke
`oQ2`	3.10 GiB	passed
`oQ2.5`	3.21 GiB	passed
`oQ2.7`	3.37 GiB	passed
`oQ3`	3.44 GiB	passed
`oQ3.5`	3.54 GiB	passed
`oQ4`	3.87 GiB	passed
`oQ5`	4.33 GiB	passed
`oQ6`	4.76 GiB	passed
`oQ8`	5.66 GiB	passed

BUILD_MANIFEST.json has the build and smoke details.

Note

Smoke checks verify local loading and short generation only. No benchmark scores are claimed.