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

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

The SAGE-MM-Qwen3-VL-4B-SFT from allenai is a 4B-parameter vision-language model fine-tuned from Qwen/Qwen3-VL-4B-Instruct, serving as the core decision-maker in the SAGE (Smart Any-Horizon Agent) system for long video reasoning, operating in two stages: Stage-1 analyzes initial sampled frames and metadata to classify queries as single-turn (immediate answer) or multi-turn (tool-required), while Stage-2 iteratively calls tools, processes observations, and refines context until resolution. It generates JSON-formatted actions for specialized tools like web-search (external knowledge), transcribe-speech (ASR on timestamps), ground-event (temporal localization), extract-video-parts (high-res frames/subclips), and analyze (detailed visual breakdown), enabling robust handling of extended videos beyond fixed horizons via dynamic tool orchestration. Released under Apache 2.0 for research/educational use per Ai2 guidelines, it requires the SAGE GitHub runtime for tool execution/parsing and supports GGUF quantizations for efficient deployment, excelling in video Q&A across sports, narratives, and events.

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