What are the hardware requirements for RefCaptioner?

RefCaptioner requires approximately 17.5 GB of VRAM. Quantized variants may run on less VRAM; see the Quantized Models section on this page.

Who developed RefCaptioner and how large is it?

RefCaptioner is developed by TengfeiLiuCoder, a model with 8b parameters. The model is published as open weights on Hugging Face and indexed on LLM Explorer with full benchmark history.

Where can I download or evaluate RefCaptioner?

RefCaptioner is hosted on Hugging Face and linked from this page. LLM Explorer also lists quantized variants and similar alternatives if available.

RefCaptioner by TengfeiLiuCoder — VRAM 17.5GB

Name: RefCaptioner
Author: TengfeiLiuCoder

RefCaptioner is an open-source language model by TengfeiLiuCoder. Features: 8b LLM, VRAM: 17.5GB, License: apache-2.0, Instruction-Based, LLM Explorer Score: 0.27.

Arxiv:2607.28509 Base model:finetune:qwen/qwen3... Base model:qwen/qwen3-vl-8b-in... Conversational En Endpoints compatible Grpo Image-text-to-text Instruct Multi-image Qwen3-vl Qwen3 vl Reference-grounding Region:us Safetensors Sharded Tensorflow Video Video-captioning

Model Card on HF 🤗: https://huggingface.co/TengfeiLiuCoder/RefCaptioner

RefCaptioner Parameters and Internals

LLM Name	RefCaptioner
Repository 🤗	https://huggingface.co/TengfeiLiuCoder/RefCaptioner
Base Model(s)	Qwen3 VL 8B Instruct Qwen/Qwen3-VL-8B-Instruct
Model Size	8b
Required VRAM	17.5 GB
Updated	2026-08-11
Maintainer	TengfeiLiuCoder
Model Type	qwen3_vl
Instruction-Based	Yes
Model Files	5.0 GB: 1-of-4 5.0 GB: 2-of-4 4.9 GB: 3-of-4 2.6 GB: 4-of-4
Supported Languages	en
Model Architecture	Qwen3VLForConditionalGeneration
License	apache-2.0
Model Max Length	262144
Transformers Version	4.57.6
Tokenizer Class	Qwen2Tokenizer
Padding Token	<\|endoftext\|>
Errors	replace

Rank the RefCaptioner capabilities

Have you tried this model? Rate its performance — this feedback helps the ML community find the right model for their needs.

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

Best Alternatives to RefCaptioner

Best Alternatives	Context / RAM	Downloads	Likes
Qwen3 VL 8B Instruct	0K / 17.5 GB	4494329	1037
Qwen3 VL 8B Instruct FP8	0K / 10.6 GB	3108604	77
Qwen3 VL 8B Heretic 1.3.0	0K / 17.5 GB	4653	12
Zhimu 1.0	0K / 18.6 GB	29	1
ZwZ 8B	0K / 17.6 GB	2586	48
... Simple Path Dense S100 Step90	0K / 17.5 GB	13	1
Qwen3 VL 8B Instruct	0K / 17.5 GB	5428	10
Qwen3 VL 8B Instruct FP8	0K / 10.6 GB	1067	7
Qwen3 VL 8B Instruct Bf16	0K / 17.6 GB	127	3
...en3 VL 8B Instruct Abliterated	0K / 17.5 GB	18299	195

Note: green Score (e.g. "73.2") means that the model is better than TengfeiLiuCoder/RefCaptioner.