SAGE-MM-Qwen3-VL-4B-SFT_RL-GGUF

SAGE-MM-Qwen3-VL-4B-SFT_RL-GGUF

The SAGE-MM-Qwen3-VL-4B-SFT_RL from allenai is a 4B-parameter reinforcement learning (RL) post-trained vision-language model, further refined from the SAGE-MM-Qwen3-VL-4B-SFT base via RL on top of Qwen/Qwen3-VL-4B-Instruct, serving as the core decision-maker in the SAGE (Smart Any-Horizon Agent) system for long video reasoning with improved performance (e.g., +4.1% over SFT per experiments) through two-stage operation: Stage-1 analyzes initial frames/metadata to classify queries as single-turn or multi-turn, while Stage-2 iteratively invokes JSON-formatted tools like web-search, transcribe-speech (ASR), ground-event (temporal localization), extract-video-parts (frames/subclips), and analyze for dynamic context updates until resolution. Designed for extended video Q&A beyond fixed horizons—handling sports, narratives, events—it requires the SAGE GitHub runtime for tool parsing/execution and outputs action strings for robust, any-length video understanding under Apache 2.0 for research/educational use per Ai2 guidelines, with GGUF quantizations available.

SAGE-MM-Qwen3-VL-4B-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):