OneStreamer

OneStreamer is a streaming video model that unifies perception, memory, and proactive response in video interaction.

OneStreamer is a streaming video large language model designed to perceive the present, remember the past, and respond at the right time. Built on Qwen3-VL-4B-Instruct, the model preserves time-grounded factual memory before future questions are known and decides when enough evidence is available to respond. It uses Proactive Hierarchical Caption Memory to retain local details and completed-event summaries alongside a recent visual window. It also employs Proactive State Transition Learning to preserve output anchors and select representative state changes. The project includes OneStreamer-1M, a dataset combining streaming captions, streaming QA, and open-source data into over one million interaction records.

OneStreamer at a glance

Best for: Researcher, Software engineer
Runs on: API, self-hosted, open source
Pricing: Free
Released: September 2026
Category: AI video tool
Website: github.com/mcg-nju/onestreamer

Key features

Performs proactive streaming video inference
Retains local details and completed-event summaries
Preserves time-grounded factual memory
Selects representative state changes and persistence
Generates control states and text from context
Supports evaluation across eight benchmarks

Use it for

Running proactive inference on streaming videos
Evaluating streaming video interaction models
Managing hierarchical caption memory for video
Generating timely responses based on visual evidence

Alternatives to OneStreamer

FreeVideo — Runs MiniMax H3 locally on consumer GPUs with as little as 8 GB of VRAM.
ATAS — Maintains persistent spatial and event memory across shots.
OpenVLHarness — Lifts six multimodal model backbones by 9.7 to 12.6 points across 22 benchmarks.
Claude Real Video — Analyzes videos locally to make visual content readable by any LLM.
TwelveLabs — It indexes video content for natural language search without manual metadata tags.
video2ctx — Provides video context with timestamped citations and uses 95% fewer tokens than transcripts.
FastVideo — Provides end-to-end post-training and real-time inference for accelerated video generation.
Self Gradient Forcing Plus — It enables continuous video generation for up to 24 hours from 5-second training windows.
H3.c — Runs native MiniMax-H3 video and audio generation locally on Apple Silicon.
Bernini — Combines a semantic planner and DiT renderer for video generation and editing tasks.
TalkLikeYou Streaming RealTime Demo — Generates 30+ FPS real-time talking-head video with low VRAM usage.
MAGI-1 — Generates synchronized audio and visuals with high-fidelity camera control.

All OneStreamer alternatives