uni-mumer-qwen3-vl-2b-GGUF

Uni-MuMER-Qwen3-VL-2B GGUF

GGUF conversions of phxember/Uni-MuMER-Qwen3-VL-2B for CrispEmbed inference.

Handwritten Mathematical Expression Recognition (HMER) model. Converts images of handwritten math into LaTeX. Fine-tuned from Qwen3-VL-2B-Instruct using multi-task training (recognition + symbol counting + position identification).

Based on the Uni-MuMER paper (NeurIPS 2025 Spotlight).

Update 2026-09-29: all files reconverted, please re-download

The previous F16 file had three zeroed tensors in LLM layer 18 (ffn_norm, attn_q_norm, attn_k_norm), and every quant was derived from it, so all of them decoded formulas wrongly. Every file has now been rebuilt from the upstream checkpoint on CI. Each LLM tensor was checked against upstream (all 308 within F16 rounding), and the output was compared end to end with transformers: the LaTeX matches exactly on the test formulas. This needs CrispEmbed main from 2026-09-29 or later, which also fixes Qwen3-VL image normalisation, deepstack placement and the chat template, and decodes Uni-MuMER with plain greedy like upstream.

Model variants

Architecture

• Vision: Qwen3-VL ViT (24L, 1024d, 16 heads, patch=16, learned pos embed)
• DeepStack: Multi-layer feature concat at layers [5, 11, 17]
• Merger: LayerNorm + GELU MLP (4096d intermediate)
• LLM: Qwen3 decoder (28L, 2048d, GQA 16/8, SwiGLU, interleaved mRoPE)
• Parameters: 2.1B
• mRoPE sections: [24, 20, 20]

Usage

# CLI — auto-detects math OCR prompt from model name
./crispembed -m uni-mumer-qwen3-vl-2b-q4_k.gguf --ocr equation.png

# Server
./crispembed-server --ocr uni-mumer-qwen3-vl-2b-q4_k.gguf --port 8080
curl -X POST http://localhost:8080/math/ocr -F "[email protected]"

from crispembed import CrispMathOcr

ocr = CrispMathOcr("uni-mumer-qwen3-vl-2b-q4_k.gguf")
latex = ocr.recognize("equation.png")
print(latex)  # x ^ { 2 } + 2 x y + y ^ { 2 } = 0

Accuracy

• ~82% ExpRate on CROHME (handwritten math benchmark)
• Replaces CC BY-NC-SA models (PosFormer/BTTR/HMER at 57%) with Apache-2.0 licensed alternative

License

Apache-2.0 — same as the base model.

Credits

Original model by BFlameSwift/Uni-MuMER (phxember on HuggingFace). GGUF conversion and inference engine by CrispEmbed.

Provenance and EU AI Act Art. 53 note

• Upstream model: phxember/Uni-MuMER-Qwen3-VL-2B — published by phxember.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.