MiniMax-H3-Prompt-Rewriter-Qwen3.5-9B-Ref2VA-GGUF

GGUF conversions of lightx2v/MiniMax-H3-Prompt-Rewriter-Qwen3.5-9B-Ref2VA. The model rewrites user requests into structured MiniMax-H3 video and audio prompts. These are complete model weights with a separate vision projector. Load the language file with -m and the projector with --mmproj.

Tasks: Ref2VA with ordered image references, sampled video frames and explicit audio descriptions.

Output: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music.

Files

Sizes are decimal GB of files on disk. Runtime memory also includes the context, vision processing and backend allocations. Both quantizations use the same F16 projector extracted from this checkpoint.

conversion.json records source/output SHA-256 hashes and the conversion tools. The BF16 control used for validation is not required to run these files.

ComfyUI

Install MiniMax-H3 Prompt Rewriter and add MiniMax-H3 Qwen Rewriter with Qwen Rewriter Options. Choose this checkpoint in Model list, or put the model and projector together under ComfyUI/models/LLM/MiniMax-H3-Prompt-Rewriter-Qwen3.5-9B-Ref2VA-GGUF/ for local discovery.

The node provides input checkboxes, the ordered reference strip, + instr, aspect ratio, duration, bypass/repeat and the prompt library. Its references output carries the original enabled conditioning to Reference Slots or Reference Encode. Select a task supported by this checkpoint.

llama.cpp

Use a build that supports qwen35 and its Qwen3-VL vision projector. Conversion used llama.cpp commit cb26014d965036f0eacdab7387b222d28d36b9e6 (b10310).

hf download pytraveler/MiniMax-H3-Prompt-Rewriter-Qwen3.5-9B-Ref2VA-GGUF --include '*Q5_K_M.gguf' 'mmproj-*.gguf' --local-dir .
llama-server -m MiniMax-H3-Prompt-Rewriter-Qwen3.5-9B-Ref2VA-Q5_K_M.gguf \
  --mmproj mmproj-MiniMax-H3-Prompt-Rewriter-Qwen3.5-9B-Ref2VA-F16.gguf --jinja \
  --ctx-size 32768 --n-gpu-layers 999 --host 127.0.0.1 --port 8081

Supply the matching official writing guide and task request, as in the upstream inference example. Image parts must stay interleaved with their labels. Video references are sampled images with source timestamps. The ComfyUI node builds this request automatically.

The upstream defaults are temperature=0.7, top_p=0.8, top_k=20 and max_tokens=16384, with thinking enabled. A smaller budget can be used with thinking disabled through chat_template_kwargs.enable_thinking=false.

Audio: this checkpoint has no audio encoder. Provide an explicit description of the reference sound. The original Ref2VA helper labels audio but does not send its waveform; these conversions do not add a hearing capability.

Validation

Q5_K_M and Q4_K_M loaded on an NVIDIA GeForce RTX 5090 and returned the task sections from image references. The original SFT also completed a smoke test (BF16 for 9B, NF4 for 27B). Exact settings, timings and H3 findings are recorded in validation.json.

Quantization changes the weights and can change the wording. A successful smoke test checks loading, media input and output structure; it is not a quality score.

Conversion and source

Source: lightx2v/MiniMax-H3-Prompt-Rewriter-Qwen3.5-9B-Ref2VA at 4b602d004317e6c7d6a9dc113bb3280835b49b71. The official convert_hf_to_gguf.py produced a BF16 language-model control with --no-nextn and an F16 vision projector with --mmproj from the same SFT directory. llama-quantize produced Q5_K_M and Q4_K_M language files. The vision projector was kept at F16; no additional LoRA is needed.

The writing guides are fetched from their original source by the node and are not redistributed in this repository.

Licence and credits

The original checkpoint includes an Apache 2.0 licence, copied here as LICENSE. Credit for the fine-tuned model belongs to LightX2V and for its base model to Qwen.

Prompt format and video generation: MiniMax-H3.

Downloads last month
1,468
GGUF
Model size
9B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

5-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for pytraveler/MiniMax-H3-Prompt-Rewriter-Qwen3.5-9B-Ref2VA-GGUF

Quantized
(3)
this model