# Publish Check: Qwen3.5-2B-oQ8-mtp

## Model

- Published repo: [`dylpickle13/Qwen3.5-2B-oQ8-mtp`](https://huggingface.co/dylpickle13/Qwen3.5-2B-oQ8-mtp)
- Base model: [`Qwen/Qwen3.5-2B`](https://huggingface.co/Qwen/Qwen3.5-2B)
- Base revision: `15852e8c16360a2fea060d615a32b45270f8a8fc`
- License: Apache-2.0
- Runtime target: oMLX / MLX
- Quantization/conversion: oMLX/oQ level 8, affine 8-bit, group size 64, bfloat16 preserved tensors, `text_only=false`, `preserve_mtp=true`

## Local artifact verification

- Local source path: `/Users/dylanrapanan/.omlx/models/Qwen3.5-2B-oQ8-mtp`
- Required files present: PASS after adding upstream `LICENSE`
- `model.safetensors` size: `2,735,417,896` bytes
- safetensors metadata: `{"format": "mlx"}`
- tensor count: `1020`
- MTP tensor count: `29`
- vision tensor count: `297`
- oMLX discovery before upload: `model_type=vlm`, `mtp_compatible=true`

## Local oMLX smoke tests

- oMLX server/version: local oMLX server; `mlx-lm 0.31.3`, `mlx-vlm 0.5.0`
- loaded model name: `Qwen3.5-2B-oQ8-mtp`
- text smoke test: PASS — prompt `Reply with exactly OK.` returned `OK`
- image smoke test: PASS — generated red square returned `red`
- notes: low-latency camera profile applied with 8192 context, 512 max tokens, thinking disabled, `mtp_enabled=true`, `vlm_mtp_enabled=false`

## Private Hugging Face upload

- private repo created: PASS
- uploaded repo: [`dylpickle13/Qwen3.5-2B-oQ8-mtp`](https://huggingface.co/dylpickle13/Qwen3.5-2B-oQ8-mtp)
- initial private upload commit: `e5394295ebc0195d3944cc416d56c7a1d8de4b3d`
- license-added private commit: `65e35d8d606f85dd1c1f45f88b2035ab7ea7a474`
- uploaded files:
  - `.gitattributes`
  - `README.md`
  - `LICENSE`
  - `chat_template.jinja`
  - `config.json`
  - `merges.txt`
  - `model.safetensors`
  - `preprocessor_config.json`
  - `tokenizer.json`
  - `tokenizer_config.json`
  - `vocab.json`

## Clean-download verification

- clean download path: `projects/omlx/quant-publisher/downloads/Qwen3.5-2B-oQ8-mtp-hf-clean`
- file list verified: PASS
- safetensors opens: PASS
- tensor counts match: PASS
- clean download tensor count: `1020`
- clean download MTP tensor count: `29`
- clean download vision tensor count: `297`
- oMLX clean load: PASS — local oMLX copy was deleted, then re-downloaded from the private Hugging Face repo, discovered, loaded, and tested
- text smoke test: PASS — `OK`
- image smoke test: PASS — `red`

## Public release

- repo made public: PASS
- anonymous access verified: PASS — public API and raw README returned HTTP 200
- final public URL: <https://huggingface.co/dylpickle13/Qwen3.5-2B-oQ8-mtp>
- public release commit verified: `18207548b8ce5bc8d699d35cbe9dcf73d0bbd056`
- final verification note: this `PUBLISH_CHECK.md` was updated after public release, so the repository latest commit may be newer than the public release commit above.

## Backup

- Google Drive backup: NA for this run unless requested separately
- backup folder URL: NA
- checksum notes: safetensors metadata and tensor counts recorded above

## Cleanup

- old local `Qwen3.5-0.8B-oQ8-mtp` removed from oMLX: PASS
- temporary base source `Qwen/Qwen3.5-2B` removed from oMLX: PASS
- clean download removed: PASS
- local project copy removed or retained intentionally: weights kept only under ignored `downloads/` temporarily for verification; no weights committed