Qwen3-VL-8B-CrossImage-SFT-GGUF

Qwen3-VL-8B-CrossImage-SFT is a multi-image reasoning model built on Qwen3-VL-8B-Instruct. It was fine-tuned on the 36,495-trace CrossImage-CoT-36K dataset, then weight-interpolated with the base model (WiSE-FT, 0.3 SFT plus 0.7 base). The chain-of-thought traces came from GPT teachers working without the gold answer, and a trace was kept only if the teacher's first answer was correct. Training was full fine-tuning of the language model and vision-language merger with the vision encoder frozen, for 1 epoch at learning rate 1e-5. The interpolation weight was picked on a 25% development split because the SFT-only model improved multi-image skills but weakened others, such as abstaining on unanswerable items and some single-image perception. The model writes a short structured rationale (Understand, Perceive one line per image, Connect, Conclude) before a final <answer> tag, and it works best with the card's exact system prompt and instruction format. On the held-out 75% of items it beats the base model on BLINK (67.6 vs 63.7) and MIRB (59.5 vs 54.2), both significant. It is marginally ahead on MuirBench (76.1 vs 75.1) and tied on Mantis-Eval (80.4), and the card says neither of those differences is statistically significant. Limitations the authors list include weaker abstention than the base model on MuirBench's unanswerable image-option questions, and perception errors inherited from the teacher traces, with about 15% of training traces flagged for at least one wrong per-image statement. The card says it is for research use only, and the weights are released under Apache-2.0.

Model Files

File Name Quant Type File Size File Link Description
Qwen3-VL-8B-CrossImage-SFT.BF16.gguf BF16 16.4 GB Link Full BF16 weights. Highest quality, largest file size.
Qwen3-VL-8B-CrossImage-SFT.Q3_K_L.gguf Q3_K_L 4.43 GB Link Lower quality but usable, good for low RAM availability.
Qwen3-VL-8B-CrossImage-SFT.Q3_K_M.gguf Q3_K_M 4.12 GB Link Low quality.
Qwen3-VL-8B-CrossImage-SFT.Q4_K_M.gguf Q4_K_M 5.03 GB Link Good quality, default size for most use cases, recommended.
Qwen3-VL-8B-CrossImage-SFT.Q4_K_S.gguf Q4_K_S 4.8 GB Link Slightly lower quality with more space savings, recommended.
Qwen3-VL-8B-CrossImage-SFT.Q5_K_M.gguf Q5_K_M 5.85 GB Link High quality, recommended.
Qwen3-VL-8B-CrossImage-SFT.Q5_K_S.gguf Q5_K_S 5.72 GB Link High quality, recommended.
Qwen3-VL-8B-CrossImage-SFT.Q8_0.gguf Q8_0 8.71 GB Link Extremely high quality, generally unneeded but max available quant.
Qwen3-VL-8B-CrossImage-SFT.mmproj-bf16.gguf mmproj-bf16 1.16 GB Link Multimodal projection file in BF16 format. Used for vision/language models.

llama.cpp

LLM inference in C/C++ β€” https://github.com/ggml-org/llama.cpp

Downloads last month
464
GGUF
Model size
8B params
Architecture
qwen3vl
Hardware compatibility
Log In to add your hardware

3-bit

4-bit

5-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for prithivMLmods/Qwen3-VL-8B-CrossImage-SFT-GGUF

Quantized
(1)
this model

Dataset used to train prithivMLmods/Qwen3-VL-8B-CrossImage-SFT-GGUF

Collection including prithivMLmods/Qwen3-VL-8B-CrossImage-SFT-GGUF