SAGE-MM-Qwen3-VL-8B-SFT_RL-GGUF

SAGE-MM-Qwen3-VL-8B-SFT_RL-GGUF

The SAGE-MM-Qwen3-VL-8B-SFT_RL from allenai is an 8B-parameter reinforcement learning (RL) post-trained vision-language model, refined via RL from the SAGE-MM-Qwen3-VL-8B-SFT (fine-tuned from Qwen/Qwen3-VL-8B-Instruct), serving as the core decision-maker in the SAGE (Smart Any-Horizon Agent) system for long video reasoning with superior performance gains over SFT through two-stage operation: Stage-1 analyzes initial sampled frames/metadata to classify queries as single-turn (immediate answers) or multi-turn (tool-required), while Stage-2 iteratively generates JSON-formatted tool calls for web-search (external knowledge), transcribe-speech (timestamped ASR), ground-event (temporal localization), extract-video-parts (high-res frames/subclips), and analyze (visual breakdown) to progressively update context until resolution. Engineered for arbitrary-length video Q&A across sports, narratives, events, or timelines beyond fixed horizons, it requires the SAGE GitHub runtime for tool parsing/execution/observation feedback, achieving state-of-the-art on benchmarks like MINERVA under Apache 2.0 for research/educational use per Ai2 guidelines, with GGUF quantizations (Q4_K_S/M recommended) for efficient deployment.

SAGE-MM-Qwen3-VL-8B-SFT_RL [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):