GitHub repository
View the README, examples, and source code.
mJev-Qwen3-VL-4B-RLCD
Explore and download the model fine-tuned from Qwen3-VL-4B-Instruct.
Six-second video example
The project usedQwen3-VL-4B-Instruct with the Transformers backend to answer three questions about a six-second animation. The commands below use the same model revision, inputs, and sampling settings as the published run.
Original video · Full output
Prepare your environment
Prepare Linux, Python 3.11 or later, a compatible NVIDIA driver, and system FFmpeg. The project’s published inference record for the six-second video used one 24 GB NVIDIA GPU.torchcodec decodes video, and PyTorch runs inference on the GPU.
Run the video example
Run this command from the repository root. The example JSON contains the video path, sampling settings, and three questions. The command uses GPU0, enables prefix caching, and processes the three questions in a batch.
outputs/motion-demo.json to inspect the results. Video example input
To run the same video with the fine-tuned mJev model, download SoMarkAI/mJev-Qwen3-VL-4B-RLCD and point --model to its local directory. Set a different --output path so you can compare the two results. Fine-tuned model download and quickstart
Use your own media
Save the following JSON astask.json, place clip.mp4 in the same directory, and point --input to task.json. media_path is relative to the input JSON file. For reproducible video results, set the sampling rate and the maximum number of frames.
"modality": "image", point media_path to the image, and remove video_options. Use audio for audio alone or audio_video to analyze video and its audio track together. Both audio modes use the Qwen3-Omni model. See the Transformers deployment guide for dependencies and startup instructions.
Read the results
The output is a JSON array with one entry per question.decision contains the selected candidate. candidates lists every option with its raw_logit and probability. The probabilities are normalized over the supplied candidates for that question, and probability_sum reports their total.
Candidate answers must be distinct, nonempty strings. The Transformers demo limits the full input to 4,000 tokens by default and returns an error if it exceeds the limit. To check the input length and candidate labels before loading the model weights, add --check-only and omit --output to see the results in your terminal.
When scores tie, mJev selects the earlier candidate in input order.
Models and inputs
Both models use the same format for questions and candidates.
- Compare the two model families, dependencies, and supported inputs. Model guide
- For vLLM, follow the project’s pinned version and GPU configuration. vLLM deployment guide
Public evaluation
- Accuracy. On the same 195 held-out questions from
mJev-Compositional-VQA, the base model answered 152 correctly (77.95%) and the fine-tuned mJev model answered 156 correctly (80.00%). Model card and results - Prefix caching. On one 24 GB NVIDIA GPU, the project tested 16 questions about the same video. Caching reduced total latency from 32.62 s to 6.29 s and increased peak allocated GPU memory from 15.80 GiB to 20.60 GiB. The project also recorded results for shorter inputs and single questions. Cache benchmark conditions

