Mage VL

Mage-VL is a codec-native, proactive-streaming multimodal foundation model for image and video understanding, whose visual encoder is trained entirely from scratch at a compact 4B scale. It targets a modern Moravec's paradox of VLMs — strong at complex offline reasoning, yet slow and compute-heavy on simple real-time streaming perception. Instead of decoding video into uniformly-sampled frames and pushing a dense grid of patch tokens through a frozen web-pretrained ViT, Mage-VL follows the structure of modern video codecs: it separates a stream into anchor (I) frames and predicted (P) frames, keeps every anchor patch, and retains only the predicted-frame patches where the codec spends bits — the regions carrying real motion and new detail. This codec-aligned sparsity cuts visual tokens by over 75% while preserving spatio-temporal context, yielding up to 3.5× wall-clock inference speedup over uniform frame sampling.

The system pairs two components:

• Mage-ViT — a from-scratch Codec-ViT visual encoder that allocates tokens by codec-derived spatio-temporal importance, on a shared 16×16 patch grid with 3D rotary position encoding. It is codec-agnostic: the same interface accepts a traditional codec (H.264/AVC, HEVC/H.265) via motion vectors + residual energy, or a neural codec (DCVC-RT) via its learned rate map — no architecture or retraining change.
• Qwen3-4B causal decoder — a Qwen3-4B-Instruct-2507 language backbone (the only pretrained component) that consumes Mage-ViT's variable-length token stream through a lightweight two-layer MLP projector, with a unified interface for images, short/long/ultra-long video, and streaming.

On top of this pair, a System 1 & System 2 dual-process design adds proactive streaming inside a single model: a lightweight cognition gate (System 1) watches each rolling codec window and stays silent on routine content, invoking the full VLM (System 2) only when a response-worthy event completes — no multi-agent pipeline required.

✨ Highlights

• Codec-native & from scratch. The entire visual stack is trained from scratch — no billion-scale image-text ViT initialization. The bio-inspired predictive-patch mechanism (I/P frames at 16×16) cuts visual-token consumption by over 75% (~1/8 or less of dense frame sampling), letting the model train on videos 8× longer under the same budget.
• Codec-native speedup. Codec tokenization sets a superior accuracy–efficiency frontier — up to 3.5× wall-clock inference speedup over uniform frame sampling at matched accuracy, and the fastest of all compared models on most video benchmarks (single 8×B200 node).
• Native-resolution scaling. Variable-resolution pretraining lets Mage-ViT improve monotonically with the token budget (peaking >96.1% Food-101 / >86.3% ImageNet at 676 tokens) where fixed-resolution encoders saturate or degrade.
• Matched-LLM video gains. With the 4B Qwen3 backbone held fixed and only the ViT swapped, Mage-VL improves over Qwen3-VL-4B on every reported video and temporal-grounding benchmark — largest on localization-heavy tasks (+22.5 QVHighlight, +17.1 ActivityNet, +11.0 VSI-Bench, +24.5 VideoEval-Pro).
• Strong for its size. On par with Qwen3-VL-4B on static images, and clearly ahead on video understanding and spatial intelligence (+11.0 VSI-Bench, +53.1 CrossPoint, +5.2 EmbSpatial, +22.5 QVHighlight).
• Proactive streaming, single model. A frozen-backbone cognition gate delivers low-latency, event-gated commentary; it tops TimVal / F1 / ROC-AUC / PR-AUC on SoccerNet streaming and generalizes to real 2026 World Cup broadcasts.

📥 Model

A single checkpoint, microsoft/Mage-VL, is one unified model that simultaneously provides image & video understanding and the proactive streaming gate — the same weights answer offline image/video questions and drive event-gated commentary. It covers every Mage-VL capability: image understanding, frame-sampled video, traditional H.264/HEVC codec video, neural DCVC-RT codec video, and event-gated streaming. The repository bundles the codec processor, the neural codec package, and the proactive gate weights — no separate understanding, NVC, or streaming checkpoint is required.

We additionally release microsoft/Mage-ViT — the standalone visual encoder from the two-stage, from-scratch ViT pre-training. This is the ViT-pre-trained checkpoint only: it has not gone through the joint VLM training with the language model. Use it as a data-efficient, codec-native visual encoder or as a drop-in ViT for your own multimodal training.

🏗️ Architecture

Mage-ViT — a from-scratch Codec-ViT visual encoder. On a 16×16 patch grid it keeps all anchor (I) frame patches and only the motion-salient predicted (P) frame patches, cutting visual tokens by over 75% while a shared 3D RoPE preserves spatio-temporal positions.

Mage-VL — a unified model where projected visual tokens and text tokens share one causal Qwen3 decoder. Still images become a single spatial block; videos become temporally-ordered codec windows. In streaming mode, a lightweight cognition gate predicts p_speak = g(h_t) per rolling window (over a recurrent streaming memory kept by an event-preserving feature extractor) and triggers generation when p_speak ≥ τ; the response is decoded by the frozen base model from a local sliding window of the most recent codec segments, and a text query can be injected at any time.

Training — a progressive five-stage supervised curriculum (no preference/RL post-training) that produces one unified model:

1. Multimodal alignment via captions — ~350M dense image captions + 4.2M short-video captions.
2. Instruction tuning + short temporal grounding — ~54M image-instruction samples + 3.4M 30–180s video captions.
3. Temporal-horizon expansion — medium/long video (LLaVA-Video, TimeLens, VideoChat-Flash, Molmo2) with retained image SFT.
4. Codec-native long-context adaptation — 350K long videos as rolling codec windows (up to 384/768 frames).
5. Proactive streaming alignment — a cognition gate fine-tuned on ~3.3M streaming samples with the visual encoder and LLM kept frozen (only the gate is trained).

The five stages together produce a single unified model, Mage-VL, that handles image understanding, offline video reasoning, and proactive streaming — no separate variants are shipped.

Two parts of the pipeline apply an AI4AI (AI-for-AI) paradigm: (1) dense recaptioning runs through an agentic closed loop where a GPT-5 rubric scorer grades captions and a Copilot coding agent co-designs the prompt and harness code (e.g. rendering timestamp overlays) under a human validation gate — improving every downstream OCR/doc/chart/perception benchmark and inspiring SkillOpt-Lite; and (2) Stage-3 uses AI-based diagnostics to decide which video categories, resolutions, and frame counts to train on.

📊 Performance

Performance comparison across models. Mage-VL-4B and Qwen3-VL-4B use the same 4B Qwen3 LLM backbone; Phi-4-Multimodal-Instruct (Phi-4-MM, 5.6B) and Phi-4-Reasoning-Vision (Phi-4-R-V, 15B) are reported for reference. – = not run. Bold = best in row.

Bold = best in row.

SoccerNet — response timing (StreamMind protocol, codec-native inputs, zero-tolerance canvas matching). Bold = best in column.

JoyAI's high TriggerAcc comes from predicting silence almost everywhere under SoccerNet's heavy class imbalance, so it collapses on the precision-sensitive metrics; StreamMind is trained in-distribution on SoccerNet, whereas Mage-VL is not.

OVO-Bench — online video understanding (SimpleStream recent-window protocol, 4 frames @ 1 fps; no streaming-specific fine-tuning). Mage-VL sets a new state-of-the-art overall score among streaming architectures. RT-Avg / BT-Avg are the Real-Time Visual Perception / Backward Tracing sub-task averages; Overall is their mean. Bold = best model per column (Human is the reference upper bound).

† HERMES = Qwen2.5-VL-7B + HERMES (4K tokens). Baseline results and table structure follow SimpleStream.

🔬 Key Findings

Beyond the model, the report distills seven empirical findings for efficient multimodal training:

1. Data-efficient tokenizer. Mage-ViT trains on only 560M unlabeled images + 100M video frames, yet matches SigLIP2 pretrained on billions of image-text pairs.
2. Variable-resolution pretraining scales monotonically. Quality keeps improving with the visual-token budget instead of saturating/degrading like fixed-resolution encoders.
3. Codec-native tokenization sets a better accuracy–efficiency frontier — up to 3.5× wall-clock inference speedup over uniform frame sampling.
4. Explicit VideoQA SFT is redundant. Dense video captions + standard image SFT are sufficient for strong zero-shot VideoQA.
5. Motion–spatial synergy. Dynamic video training substantially improves static 2D/3D spatial reasoning.
6. AI4AI data pipeline. Agentic closed-loop feedback + prompt/code co-design systematically lift caption quality and downstream scores (inspired SkillOpt-Lite).
7. Zero-Vision SFT for multimodal RL. Bypassing visual SFT in favor of pure-text reasoning SFT unlocks stronger multimodal RL — a compute-efficient path.

🚀 Quick Start

A single checkpoint, microsoft/Mage-VL, covers every capability below.

Installation

For offline Transformers inference:

pip install "transformers>=5.7" accelerate pillow torch torchvision \
  opencv-python codec-video-prep

Codec-based video inference also requires ffmpeg and ffprobe on PATH.

Using 🤗 Transformers directly

Mage-VL ships its modeling, processing, and chat-template code inside this repository, so image and frame-sampled video inference need no script download and no extra package beyond transformers:

import torch
from PIL import Image
from transformers import AutoModelForCausalLM, AutoProcessor

model_id = "microsoft/Mage-VL"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, torch_dtype="auto", device_map="auto"
).eval()

messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "Describe this image in detail."},
]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = processor(
    text=[text], images=[Image.open("examples/dog.jpg").convert("RGB")], return_tensors="pt"
)
inputs = {k: (v.to(model.device) if hasattr(v, "to") else v) for k, v in inputs.items()}
inputs["pixel_values"] = inputs["pixel_values"].to(model.dtype)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(processor.tokenizer.decode(
    output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True
).strip())

For video, use {"type": "video"} in the message and pass a list of PIL frames as videos=[frames]:

import cv2
import numpy as np
from PIL import Image

def sample_video(path, num_frames=32):
    capture = cv2.VideoCapture(path)
    total = int(capture.get(cv2.CAP_PROP_FRAME_COUNT))
    frames = []
    for index in np.linspace(0, total - 1, min(num_frames, total), dtype=int):
        capture.set(cv2.CAP_PROP_POS_FRAMES, int(index))
        ok, frame = capture.read()
        if ok:
            frames.append(Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)))
    capture.release()
    return frames

messages = [{"role": "user", "content": [
    {"type": "video"},
    {"type": "text", "text": "Describe this video."},
]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = processor(
    text=[text],
    videos=[sample_video("examples/soccer-broadcast.mp4", 32)],
    return_tensors="pt",
    padding=True,
)

Add attn_implementation="sdpa" to from_pretrained if flash-attn is not installed. Codec-based video (H.264/HEVC, DCVC-RT) and streaming commentary use inference.py below.

Examples

Two sample inputs ship with this repository:

Offline inference

Download inference.py. Offline mode loads the checkpoint with AutoModelForCausalLM.from_pretrained and supports images, frame sampling, and both codec engines:

python inference.py --mode offline --image examples/dog.jpg \
  --question "Describe this image in detail."

python inference.py --mode offline --video examples/soccer-broadcast.mp4 \
  --video-backend frames --num-frames 32 \
  --question "Describe this video."

python inference.py --mode offline --video examples/soccer-broadcast.mp4 \
  --video-backend codec --codec-engine traditional --num-frames 32 \
  --question "Describe this video."

python inference.py --mode offline --video examples/soccer-broadcast.mp4 \
  --video-backend codec --codec-engine neural --num-frames 32 \
  --question "Describe this video."

Online inference

Online mode talks to an OpenAI-compatible SGLang server. First build and launch the server with the Mage-VL SGLang branch (building it needs protobuf-compiler and a Rust toolchain):

sudo apt-get update && sudo apt-get install -y protobuf-compiler
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs \
  | sh -s -- -y --profile minimal --default-toolchain 1.90.0
source "$HOME/.cargo/env"

git clone -b feat/mage-vl https://github.com/kcz358/sglang
cd sglang
pip install -e 'python[all]'
python -m sglang.launch_server \
  --model-path microsoft/Mage-VL \
  --trust-remote-code

Then send an image or sampled video frames to the running server:

pip install openai

python inference.py --mode online --image examples/dog.jpg \
  --question "Describe this image in detail." \
  --base-url http://localhost:30000/v1

python inference.py --mode online --video examples/soccer-broadcast.mp4 \
  --num-frames 32 \
  --question "Describe this video." \
  --base-url http://localhost:30000/v1

Use --model, --max-new-tokens, and --api-key to override their defaults.

Streaming inference

streammind_gate.safetensors in this repository holds the event gate. Streaming inference splits a video into non-overlapping segments, stays silent on routine content, and generates a caption only when a response-worthy event is detected. Run it with inference_streaming.py from the GitHub repository:

python inference_streaming.py \
  --video examples/soccer-broadcast.mp4 \
  --video_backend codec \
  --segment_sec 8

[t=0.0-8.0s] gate=silence (p=0.19)
[t=8.0-16.0s] gate=response (p=0.55) -> The video features a live sports broadcast from BBC Sport, set in a large stadium filled with spectators. The broadcast focuses on a football match between England and Argentina, with the score displayed as England 1, Argentina 2. [...]
[t=16.0-24.0s] gate=response (p=0.73) -> The video features a sports broadcast set in a large stadium filled with spectators. Four commentators are gathered around a table with a 'BBC Sport' logo, each holding a yellow microphone. [...]
[t=24.0-30.0s] gate=silence (p=0.31)

The gate is trained on codec inputs, so --video_backend codec is the intended setting. Use --video_backend frames for direct frame sampling. Additional controls include --num_frames, --cur_fps, --max_segments, --max_new_tokens, --gate_threshold, and --attn_impl.

📝 Citation

@article{mage2026magevl,
  title={Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model},
  author={Yang, Senqiao and Zhang, Kaichen and Jia, Zhaoyang and Guo, Jinghao and Shen, Yifei and Zhang, Xinjie and Zhang, Xiaoyi and Wang, Haoqing and Li, Xiao and An, Xiang and Xie, Yin and Liu, Zhening and Guo, Xun and Li, Jiahao and Zheng, Shicheng and Wang, Jinglu and Guo, Zongyu and Xie, Wenxuan and Zheng, Zihan and Luo, Yuxuan and Li, Bin and Lu, Yan},
  journal={arXiv preprint arXiv:2607.24904},
  year={2026}
}

📄 License

Mage-VL is released under the Apache-2.0 License.