MMFineReason-4B-f32-GGUF

MMFineReason-4B-f32-GGUF

MMFineReason-4B from OpenDataArena is a 4B-parameter vision-language model fine-tuned from Qwen3-VL-4B-Instruct on the massive MMFineReason dataset (1.8M high-quality samples, 5.1B solution tokens with average 2,910-token long-form Chain-of-Thought reasoning covering mathematics 79.4%, science 13.8%, puzzles/games 4.6%, and OCR/general 2.2%), using a two-stage pipeline of supervised fine-tuning (SFT) on MMFineReason-1.8M-SFT followed by reinforcement learning (RL) with GSPO on MMFineReason-1.8M-RL to achieve state-of-the-art multimodal reasoning performance that surpasses the larger Qwen3-VL-8B-Thinking (73.9 vs 72.5 average score). Trained via a systematic three-stage data-centric pipeline—large-scale collection/standardization, CoT rationale generation from Qwen3-VL-235B-A22B-Thinking, and difficulty-aware filtering revealing a "less is more" effect where just 7% (123K samples) matches full dataset performance—it excels on STEM diagrams, visual puzzles, complex charts (90.8% CharXiv, 75.6% RealWorldQA), and math benchmarks like 83.4% DynaMath while boosting general capabilities through reasoning-focused composition. Part of the MMFineReason family (2B/4B/8B scales), this parameter-efficient model demonstrates how high-quality, visually-grounded CoT data closes the gap between open-source VLMs and proprietary systems, available for deployment via standard Transformers frameworks.

MMFineReason-4B [GGUF]

Quants Usage

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):