cstr/texteller-3-GGUF

TexTeller 3.0 GGUF

GGUF conversions of OleehyO/TexTeller for CrispEmbed inference.

Lightweight math OCR model (310M params). Converts images of printed and handwritten math expressions into LaTeX. Trained on 80M image-LaTeX pairs, supports mixed English/Chinese formulas.

Model variants

File	Quant	Size	Notes
texteller-3-f16.gguf	F16	568 MB	Full precision
texteller-3-q8_0.gguf	Q8_0	302 MB	Recommended
texteller-3-q4_k.gguf	Q4_K	169 MB	Max compression

Architecture

• Encoder: ViT (12L, 768d, 12 heads, patch=16, 448x448 grayscale input)
• Decoder: TrOCR (12L, 1024d, 16 heads, 4096 FFN, GELU, cross-attention)
• Vocabulary: 15,000 tokens (BPE + 3,500 Chinese characters)
• Parameters: 310M total (86M encoder + 224M decoder)
• Input: 1-channel grayscale, 448x448 pixels
• Output: LaTeX string

Usage

# CLI
./crispembed -m texteller-3-q8_0.gguf --ocr formula.png

# Server
./crispembed-server --ocr texteller-3-q8_0.gguf --port 8080
curl -X POST http://localhost:8080/math/ocr -F "image=@formula.png"

from crispembed import CrispMathOcr

ocr = CrispMathOcr("texteller-3-q8_0.gguf")
latex = ocr.recognize("formula.png")
print(latex)  # \mathrm{x}+\mathrm{y}

License

Apache-2.0 — same as the base model.

Credits

Original model by OleehyO/TexTeller. GGUF conversion and inference engine by CrispEmbed.