Converted in-house from Qwen's official weights — deliberately not re-published from a third-party quantization, so the bundle shipped to devices has no intermediate publisher in its supply chain.


source repo	`Qwen/Qwen3.5-2B` (official)
source revision	`15852e8c16360a2fea060d615a32b45270f8a8fc`
converted with	`mlx_lm` 0.31.2, `mlx` 0.31.1
command	`mlx_lm.convert --hf-path Qwen/Qwen3.5-2B --q-bits 4 --q-group-size 64`
result	4.503 bits/weight

The conversion is reproducible: pin the source revision above, re-run the command, and compare against the SHA-256 values in manifest.json.

The vision tower is dropped during conversion (Model.sanitize discards vision_tower.* / model.visual.*), so this bundle carries language weights only — 1059 MB versus 1749 MB for the multimodal 4-bit build. Munin never uses the image path.

manifest.json carries a SHA-256 for every file. The app hashes the bundle after download and refuses to activate the model on any mismatch.

Notes for integrators

Thinking mode is on by default. Apply the chat template with enable_thinking=False to get a clean answer; otherwise the model emits its reasoning trace as the response body. The template prefills an empty <think>\n\n</think> block when thinking is disabled.


context	262 144
layers	24 (18 linear-attention, 6 full-attention)
vocab	248 320
eos	`<

Licence

Apache 2.0, inherited from the base model.

asksolz/munin-qwen35-2b-text-4bit-mlx

munin-qwen35-2b-text-4bit-mlx

Provenance

Notes for integrators

Licence

asksolz/
munin-qwen35-2b-text-4bit-mlx