# Huihui-Qwen3.6-27B-abliterated-NVFP4-TEXT-MTP — Lna-Lab variant

**Status: ✅ Slim variant created** (2026-04-26)
**Lineage**: BF16 base → NVFP4-MTP (modelopt) → **TEXT-MTP** (this slim)

## What this is

Text-only variant of `Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP` with the
vision tower stripped out. Same NVFP4 weights for the LM, same MTP head,
just no vision-language capability. Designed for pure-text workloads that
want lower VRAM and slightly faster startup.

## Differences from NVFP4-MTP (parent)

| | NVFP4-MTP (parent) | TEXT-MTP (this) |
|---|---|---|
| `architectures` | `Qwen3_5ForConditionalGeneration` | `Qwen3_5ForCausalLM` |
| `language_model_only` | `False` | `True` |
| `vision_config` | present | **removed** |
| `image_token_id`, `video_token_id` | present | **removed** |
| `vision_*_token_id` | present | **removed** |
| `model.visual.*` weights | 333 tensors (~0.92 GB BF16) | **dropped** |
| `model.embed_vision.*` weights | present | **dropped** |
| `preprocessor_config.json` | present | **removed** |
| `video_preprocessor_config.json` | present | **removed** |
| MTP head | preserved | **preserved** |
| Total file size | 20.56 GB | **19.64 GB** |

## How it was made

Re-quantization NOT performed. The existing NVFP4 weights are bit-for-bit
identical to the parent — only vision-related tensors were filtered out
during a single-pass safetensors copy.

```bash
python3 /home/tonoken/Models/slim_qwen36_27b_text_mtp.py
```

The script:
1. Reads parent `model.safetensors`, drops any tensor with `visual` /
   `vision` / `embed_vision` in the key path.
2. Writes new `model.safetensors` with the remaining 2354 / 2687 tensors.
3. Writes new `config.json` with VLM fields removed and architecture
   downgraded to `Qwen3_5ForCausalLM`.
4. Writes new `hf_quant_config.json` with vision entries pruned from the
   exclude list.
5. Copies `chat_template.jinja`, `tokenizer.*`, `generation_config.json`,
   `README.md` (skips VLM preprocessor configs).

## Use case

- Pure text inference where VLM headroom is wasted
- Slightly lower VRAM ceiling (~1 GB at load time)
- Faster startup (no vision tower init)
- Lighter Docker image footprint when the model is bundled
- Same MTP-driven decode behavior as the VLM parent

## Launch (vLLM, 1 GPU minimum)

```bash
docker run -d --name qwen36-text-mtp \
  --gpus '"device=0"' \
  --shm-size=16g \
  -v /home/tonoken/Models/huihui-ai/Huihui-Qwen3.6-27B-abliterated-NVFP4-TEXT-MTP:/models/current:ro \
  -p 8000:8000 \
  vllm/vllm-openai:cu130-nightly \
  --model /models/current \
  --served-model-name qwen36-text-mtp \
  --max-model-len 65536 \
  --gpu-memory-utilization 0.85 \
  --trust-remote-code \
  --kv-cache-dtype fp8_e4m3
```

## Provenance

- Parent: `huihui-ai/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP`
  (modelopt 0.43.0, NVFP4, MTP layer preserved as ignore)
- Base BF16: `huihui-ai/Huihui-Qwen3.6-27B-abliterated`
- Slim author: TonoKen3 / Lna-Lab (2026-04-26)