A Q6_K GGUF quantization of infly/Infinity-Parser2-Flash for llama.cpp / llama-server, so the model runs on a single consumer GPU (validated on an RTX 3080 Ti, 12 GB) without vLLM. ~4.2 GB bf16 → ~1.5 GB Q6_K weights (+ 0.67 GB f16 vision projector).

The base is a Qwen3.5-architecture vision-language model for document understanding: OCR, layout analysis, tables→HTML, charts→JSON, formulas→LaTeX, and Markdown conversion (EN/ZH).

Files

File	What
`Infinity-Parser2-Flash-Q6_K.gguf`	Q6_K-quantized weights (imatrix)
`Infinity-Parser2-Flash-mmproj-f16.gguf`	f16 multimodal projector — required for image input

Method

convert_hf_to_gguf → f16 GGUF → llama-quantize Q6_K with an importance matrix computed from a clean native-PDF document corpus (~519 k tokens). (llama-imatrix is text-only; the mmproj carries the vision tower at serve time.)

Quality (VLMEvalKit, vs published bf16)

Benchmark	bf16	Q6_K GGUF
DocVQA (val)	93.80	93.63
OCRBench	84.3	82.8
MMStar / MMBench	ref	≥ bf16

Effectively lossless for the 6-bit quant. The small OCRBench dip is not the quantization — an f16 GGUF on the same stack scores ≈ 83.0 ≈ Q6_K's 82.8, so the residual gap is the llama.cpp vision preprocessing (candle CLIP), not the 6-bit weights.

Serving (llama.cpp)

bash

llama-server \
  --model Infinity-Parser2-Flash-Q6_K.gguf \
  --mmproj Infinity-Parser2-Flash-mmproj-f16.gguf \
  --ctx-size 32768 --n-gpu-layers 99 \
  --host 0.0.0.0 --port 8105

OpenAI-compatible /v1/chat/completions with image_url content. Notes:

Reasoning-capable model: output may arrive in the reasoning_content channel (llama.cpp routes the think block there) — read it accordingly, or disable thinking.
A 16 MP page ≈ 15.6 K vision tokens, so --ctx-size 32768 comfortably fits one page + output.

spectator2026/Infinity-Parser2-Flash-GGUF

Infinity-Parser2-Flash — Q6_K GGUF (+ vision mmproj)

Files

Method

Quality (VLMEvalKit, vs published bf16)

Serving (llama.cpp)

spectator2026/
Infinity-Parser2-Flash-GGUF