OpenMOSE/Qwen3-VL-REAP-24B-A3B

OpenMOSE/Qwen3-VL-REAP-24B-A3B

Vision–Language MoE model created by applying Router-weighted Expert Activation Pruning (REAP) to Qwen3-VL-30B.

---

1. Model Summary

• Base model: Qwen3-VL-30B (vision–language MoE LLM)
• Variant name: Qwen3-VL-REAP-24B-A3B
• Architecture: Decoder-only Transformer + MoE MLP experts, with vision encoder + VL fusion as in Qwen3-VL
• Pruning method: REAP (Router-weighted Expert Activation Pruning) by Cerebras Research https://github.com/CerebrasResearch/reap
• Expert sparsity: ~25% of MoE experts pruned globally
• Active parameters: “A3B” indicates roughly ~3B active parameters per token (MoE sparse activation), while total parameters are reduced to about 24B
• Modality: Text + Vision (VL support kept intact)
• License: Apache 2.0
• Author / Maintainer: OpenMOSE
• Year: 2025

This is an unofficial community variant of Qwen3-VL, not affiliated with or endorsed by Alibaba or Cerebras Systems.

---

2. What Is REAP and What Did We Change?

REAP (Router-weighted Expert Activation Pruning) is a pruning method for MoE models that uses:

• Router statistics (routing probabilities)
• Expert activation patterns on a calibration set

to identify under-used or redundant experts and prune them while preserving model quality as much as possible.

For this model:

• We applied REAP to Qwen3-VL-30B across its MoE MLP blocks.
• ~40% of experts are pruned, based on router-weighted activation statistics.
• The routing mechanism itself is not conceptually changed; we only changed which experts remain.
• We extended the original REAP implementation to support the Qwen3-VL architecture, i.e. vision encoder + VL fusion layers, so pruning can be applied without breaking VL functionality.

In short: same REAP algorithm, adapted to Qwen3-VL, and leaving VL functionality available.

---

3. Calibration Data

The REAP pruning statistics were computed using:

• Calibration dataset: https://huggingface.co/datasets/OpenMOSE/reap-calib-mix

• This dataset is mostly synthetic, generated by Qwen3-235B-Instruct on mixed prompts designed to cover:

  • General instruction-following
  • Reasoning and long-form text

The calibration set is not used for additional fine-tuning here; it is used to measure router/expert activations to decide which experts to prune.

---

4. Why 24B-A3B? (Motivation & Hardware Footprint)

By pruning ~25% of experts and keeping VL:

• The model shrinks from ~30B total parameters to about 24B total parameters.

• With sparse MoE activation, around 3B parameters are active per token (“A3B”).

• In practice, this makes it feasible to deploy on a single 16 GB GPU:

  • No CPU offload is strictly required in Q4_K_M Quantization
  • Lower VRAM configurations may still use offload / agressive quantization.

The goal is to keep the full VL capability of Qwen3-VL-30B, while making it:

• Easier to experiment with REAP-style MoE pruning
• More practical to deploy and fine-tune on high-end but single-node hardware

---

5. Intended Use

Primary intended uses

• Research on:

  • MoE pruning and compression (especially REAP)
  • Scaling behavior of pruned MoE VL models
  • Trade-offs between expert sparsity and performance

• Experimental deployment for:

  • Vision–language assistants
  • Multimodal chatbots
  • Document + image understanding

Suitable tasks (examples)

• Multimodal chat (image + text → text)
• Image captioning / description
• Visual question answering
• General instruction-following and long-form text generation

Out-of-scope / high-risk uses

This model should not be used without additional safeguards for:

• Medical, legal, or financial advice
• Safety-critical decision making
• Political persuasion or targeted disinformation
• Any scenario where incorrect or biased outputs can cause real-world harm

---

6. Limitations & Risks

This model inherits all the limitations of Qwen3-VL-30B plus those introduced by pruning:

• Hallucinations: The model can generate plausible but incorrect facts.

• Bias & toxicity: Biases from the original training data and synthetic calibration data remain and may be amplified.

• Distribution shift from pruning:

  • Some long-tail behaviors may degrade due to pruning 40% of experts.
  • Performance may be uneven across tasks, domains, or languages not well covered in the calibration set.

• Multimodal edge cases:

  • Complex compositional visual reasoning or extremely high-resolution images may not work reliably.
  • VL behavior is preserved but not fully re-tuned after pruning.

Users should perform their own evaluation before relying on the model in any sensitive context.

---

7. How to Use

Note: Replace class names according to the transformers version you use. Here we assume Qwen3VLForConditionalGeneration and AutoProcessor are available (as in recent Qwen3-VL integrations).

import torch
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

model_id = "OpenMOSE/Qwen3-VL-REAP-24B-A3B"

processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# Example: image + text prompt
image = ...  # PIL.Image or numpy array
prompt = "Describe this image and summarize its key elements in one paragraph."

inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
with torch.no_grad():
    generated = model.generate(
        **inputs,
        max_new_tokens=512,
    )

output_text = processor.batch_decode(generated, skip_special_tokens=True)[0]
print(output_text)

For text-only usage, provide only text= to the processor.

---

8. Evaluation (Status)

• This release focuses on making the REAP-pruned VL model available.

• Quantitative benchmarks (e.g., MMBench, general QA, reasoning benchmarks) are still work in progress.

• Early qualitative checks show:

  • VL behavior is preserved after pruning.
  • Latency and memory usage are improved compared to Qwen3-VL-30B, especially on single 16 GB GPUs.

Community contributions with detailed benchmarks are very welcome.

---

9. Training & Distillation Details (High-Level)

• Base model: Qwen3-VL-30B

• Pruning method: REAP (Router-weighted Expert Activation Pruning)

• Calibration data: OpenMOSE/reap-calib-mix (mostly generated by Qwen3-235B-Instruct)

• Post-processing:

  • Router / gating structure retained
  • Experts pruned according to REAP scoring
  • No additional large-scale pretraining is performed in this release

Future versions may include post-pruning fine-tuning or distillation to recover more performance.

---

10. Community & Contribution

Let’s grow this model together as a community.

You are encouraged to:

• Run benchmarks and publish results

• Contribute scripts for:

  • Further pruning experiments
  • Quantization (e.g., GGUF, AWQ, GPTQ)
  • Long-context or domain-specific fine-tuning

• Report issues or findings about failure modes, biases, or surprising behaviors

---

11. License

• Model & code (this repository): Apache License 2.0
• The original Qwen3-VL-30B model and any downstream use must also respect their respective licenses and usage terms.

---

12. Acknowledgements

• Qwen team for building the Qwen3-VL family of models.
• Cerebras Research for the REAP method and reference implementation: https://github.com/CerebrasResearch/reap
• OpenMOSE community for experimentation, engineering, and calibration data generation.

---

2025 OpenMOSE