Reframe-GGUF

Reframe is a 9B image-text-to-text model for agentic visual reasoning, built on Qwen3.5-9B and trained on the Reframe-100K dataset. It comes from the paper "Reframe: Observation-Aligned Multi-Teacher On-Policy Distillation for Agentic Visual Reasoning" (Nanjing University). The paper's idea is to treat the observation state, not just the output distribution, as the thing to align. The two stages are described as complementary, and neither needs a task-specific SFT cold start. First, Ground-Drop-GRPO (GD-GRPO) trains spatial and temporal specialists with evidence-aware rewards on tool actions and random removal of the original visual input, so the model has to reason from what its tools return. Second, Observation-Gated MOPD (OG-MOPD) replaces deficient tool actions with canonical ones mid-rollout and resumes multi-teacher on-policy distillation from the repaired observation. Across 14 image and video benchmarks, the paper says Reframe beats both Mix-RL and vanilla MOPD. Its ablations show GD-GRPO raises observation dependence and tool gain compared with outcome-based GRPO. The model card is minimal and gives no scores, and the paper abstract gives no figures, so numbers are only in the paper. Training, inference and evaluation instructions are in the GitHub repository, and the weights are Apache 2.0.

MTP β€” Speculative Decoding: No

Model Files

File Name Quant Type File Size File Link Description
Reframe.BF16.gguf BF16 17.9 GB Link Full BF16 weights. Highest quality, largest file size.
Reframe.Q3_K_L.gguf Q3_K_L 5.03 GB Link Lower quality but usable, good for low RAM availability.
Reframe.Q3_K_M.gguf Q3_K_M 4.68 GB Link Low quality.
Reframe.Q4_K_M.gguf Q4_K_M 5.68 GB Link Good quality, default size for most use cases, recommended.
Reframe.Q4_K_S.gguf Q4_K_S 5.36 GB Link Slightly lower quality with more space savings, recommended.
Reframe.Q5_K_M.gguf Q5_K_M 6.49 GB Link High quality, recommended.
Reframe.Q5_K_S.gguf Q5_K_S 6.31 GB Link High quality, recommended.
Reframe.Q6_K.gguf Q6_K 7.36 GB Link Very high quality, near perfect, recommended.
Reframe.mmproj-bf16.gguf mmproj-bf16 922 MB 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

Quantized using llama.cpp by prithivMLmods.

Downloads last month
751
GGUF
Model size
9B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

3-bit

4-bit

5-bit

6-bit

16-bit

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

Model tree for prithivMLmods/Reframe-GGUF

Finetuned
Qwen/Qwen3.5-9B
Quantized
(1)
this model

Dataset used to train prithivMLmods/Reframe-GGUF

Collection including prithivMLmods/Reframe-GGUF