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]

File Name	Quant Type	File Size	File Link
MMFineReason-4B.IQ4_XS.gguf	IQ4_XS	2.49 GB	Download
MMFineReason-4B.Q2_K.gguf	Q2_K	1.8 GB	Download
MMFineReason-4B.Q3_K_L.gguf	Q3_K_L	2.41 GB	Download
MMFineReason-4B.Q3_K_M.gguf	Q3_K_M	2.24 GB	Download
MMFineReason-4B.Q3_K_S.gguf	Q3_K_S	2.05 GB	Download
MMFineReason-4B.Q4_K_M.gguf	Q4_K_M	2.72 GB	Download
MMFineReason-4B.Q4_K_S.gguf	Q4_K_S	2.6 GB	Download
MMFineReason-4B.Q5_K_M.gguf	Q5_K_M	3.16 GB	Download
MMFineReason-4B.Q5_K_S.gguf	Q5_K_S	3.09 GB	Download
MMFineReason-4B.Q6_K.gguf	Q6_K	3.63 GB	Download
MMFineReason-4B.Q8_0.gguf	Q8_0	4.69 GB	Download
MMFineReason-4B.f16.gguf	F16	8.83 GB	Download
MMFineReason-4B.f32.gguf	F32	17.7 GB	Download
MMFineReason-4B.i1-IQ1_M.gguf	i1-IQ1_M	1.25 GB	Download
MMFineReason-4B.i1-IQ1_S.gguf	i1-IQ1_S	1.18 GB	Download
MMFineReason-4B.i1-IQ2_M.gguf	i1-IQ2_M	1.68 GB	Download
MMFineReason-4B.i1-IQ2_S.gguf	i1-IQ2_S	1.58 GB	Download
MMFineReason-4B.i1-IQ2_XS.gguf	i1-IQ2_XS	1.48 GB	Download
MMFineReason-4B.i1-IQ2_XXS.gguf	i1-IQ2_XXS	1.37 GB	Download
MMFineReason-4B.i1-IQ3_M.gguf	i1-IQ3_M	2.13 GB	Download
MMFineReason-4B.i1-IQ3_S.gguf	i1-IQ3_S	2.07 GB	Download
MMFineReason-4B.i1-IQ3_XS.gguf	i1-IQ3_XS	1.98 GB	Download
MMFineReason-4B.i1-IQ3_XXS.gguf	i1-IQ3_XXS	1.84 GB	Download
MMFineReason-4B.i1-IQ4_NL.gguf	i1-IQ4_NL	2.6 GB	Download
MMFineReason-4B.i1-IQ4_XS.gguf	i1-IQ4_XS	2.48 GB	Download
MMFineReason-4B.i1-Q2_K.gguf	i1-Q2_K	1.8 GB	Download
MMFineReason-4B.i1-Q2_K_S.gguf	i1-Q2_K_S	1.69 GB	Download
MMFineReason-4B.i1-Q3_K_L.gguf	i1-Q3_K_L	2.41 GB	Download
MMFineReason-4B.i1-Q3_K_M.gguf	i1-Q3_K_M	2.24 GB	Download
MMFineReason-4B.i1-Q3_K_S.gguf	i1-Q3_K_S	2.05 GB	Download
MMFineReason-4B.i1-Q4_0.gguf	i1-Q4_0	2.59 GB	Download
MMFineReason-4B.i1-Q4_1.gguf	i1-Q4_1	2.84 GB	Download
MMFineReason-4B.i1-Q4_K_M.gguf	i1-Q4_K_M	2.72 GB	Download
MMFineReason-4B.i1-Q4_K_S.gguf	i1-Q4_K_S	2.6 GB	Download
MMFineReason-4B.i1-Q5_K_M.gguf	i1-Q5_K_M	3.16 GB	Download
MMFineReason-4B.i1-Q5_K_S.gguf	i1-Q5_K_S	3.09 GB	Download
MMFineReason-4B.i1-Q6_K.gguf	i1-Q6_K	3.63 GB	Download
MMFineReason-4B.imatrix.gguf	imatrix	3.87 MB	Download
MMFineReason-4B.mmproj-Q8_0.gguf	mmproj-Q8_0	454 MB	Download
MMFineReason-4B.mmproj-f16.gguf	mmproj-f16	836 MB	Download
MMFineReason-4B.mmproj-f32.gguf	mmproj-f32	1.66 GB	Download

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):