mindqtrl

qwen3vl-8b-fp8-text-only-en

Deploy Dedicated

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Get help setting up a custom Dedicated Endpoints.

Talk with our engineer to get a quote for reserved GPU instances with discounts.

README

License: apache-2.0

Modifications

Vision removed: All model.visual.* tensors (351 tensors) were dropped, leaving only the text decoder (36 layers, 903 tensors).
English-only vocab: Non-English tokens (CJK, Cyrillic, Arabic, etc.) were pruned from the tokenizer and embedding matrix. Vocab reduced from 151,936 to 105,785.
FP8 preserved: The original comfy_quant and weight_scale metadata is intact. No requantization was performed.

Base model

Original: Qwen/Qwen3-VL-8B

Files

model.safetensors — text-only weights (BF16 embeds + FP8 layer weights)
tokenizer.json — pruned BPE tokenizer
config.json — Qwen3ForCausalLM config with updated vocab_size

Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("mindqtrl/qwen3vl-8b-fp8-text-only-en")
tokenizer = AutoTokenizer.from_pretrained("mindqtrl/qwen3vl-8b-fp8-text-only-en")

Stats

Table
Metric	Value
Original size	10.59 GB
Text-only size	9.44 GB
English-only size	8.68 GB
Vocab (original)	151,936
Vocab (pruned)	105,785
Layers	36
Hidden size	4096
Attention heads	32
KV heads	8