[Model Request] Add Qwen3-VL-32B-Instruct for Vision/Document OCR

[illogicallogic4]

Request

Please add Qwen3-VL-32B-Instruct (Vision-Language Model) to Cocoon network.

Use Case: Legal Document Processing

We are building Starec - an AI legal assistant for Russian legal system:

• OCR of legal documents (court decisions, laws, case files)
• Processing scanned PDFs from pravo.gov.ru (Russian official legal portal)
• Extracting structured data from criminal case volumes
• Multi-page document analysis with 256K context

Why Qwen3-VL-32B?

Feature	Value
DocVQA accuracy	~96% (near human-level 98%)
OCRBench	88.8%
Languages	32 (including Russian/Cyrillic)
Context	256K tokens (extendable to 1M)
Architecture	Dense (more stable than MoE)

Key advantages for legal OCR:

• Robust in low light, blur, tilted text (scanned documents)
• Improved rare/ancient character recognition (legal terminology)
• Long document structure parsing (multi-page court decisions)
• JSON extraction accuracy matches GPT-4o (~75%)

Technical Details

• Model: Qwen/Qwen3-VL-32B-Instruct
• HuggingFace: https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct
• Parameters: 32B (dense)
• VRAM: ~70GB FP16, ~40GB INT8
• vLLM compatible: Yes

Alternative (smaller)

If 32B is too heavy initially:

• Qwen3-VL-8B-Instruct (~18GB VRAM) - still excellent for OCR

Contact

• Project: Starec (Legal AI Assistant)
• Wallet: UQDtVYbmmARixnYhtDQGljyoXSEtM_NoJxQ4NlwDQPuLubnG

References

• Qwen3-VL GitHub
• Qwen3-VL Technical Report
• OmniAI OCR Benchmarks

[illogicallogic4]

86 hours 35 minutes 31 seconds since submitted.

First real-world Cocoon app waiting.

@arseny30 any update?

[illogicallogic4]

Good things take time. We'll wait.