Mage-VL — Multimodal LLM AI Model
Mage-VL — 4B codec-native streaming multimodal model from Microsoft. A from-scratch Mage-ViT visual encoder on a Qwen3-4B decoder, trained to read images and video through codec structure rather than uniform frame sampling. This entry serves its image-and-text mode; its video and proactive-streaming modes are not wired up here.
Details
- CategoryMultimodal LLM
- Year0
- LicenseApache-2.0
Inputs & Outputs
- TextInputstring
Text prompt or question
- ImageInputimageoptional
Optional image to ask about. Mage-VL reads one image per request — it is a single-image-or-video model — and the text prompt is answered about it. Supports up to 4096x4096 resolution.
- TextOutputstring
Generated text response
Parameters
- max_new_tokensintegerdefault 2048range 1–8192advanced
Max new tokens (1-8192)
Maximum number of tokens the model may generate in the response. Default 2048. Lower for terse answers / cost saving; raise (up to 8192) for long-form generation. Generation also stops at natural end-of-sequence regardless of this cap.
- temperaturefloatdefault 0.7range 0–2advanced
Temperature (0-2)
Sampling temperature. 0.0 = greedy / deterministic (best for factual Q&A, OCR, code). 0.7 (default) = balanced creativity. >1.0 = more divergent / creative. Raise for brainstorming, lower for precision tasks.
- top_pfloatdefault 0.8range 0–1advanced
Top-p (nucleus)
Nucleus sampling threshold — keeps the smallest set of tokens whose cumulative probability ≥ top_p, samples from that set. 1.0 = unrestricted, 0.8-0.95 (typical defaults) = balanced. Lower values produce more focused output. Combined with top_k.
- top_kintegerdefault 20range 1–100advanced
Top-k
Top-k sampling — restricts the candidate set to the k highest-probability tokens. 50 is permissive, 20 (Qwen default) is tighter. Lower for more deterministic output, higher for variety.
Alternatives in Multimodal Language Models
- Agnes-3.0-Flash
Agnes 3.0 Flash Preview — 33B hybrid-attention vision-language model, 262K context, bf16. Open-weights preview checkpoint, distinct from the production Agnes 3.0 Flash.
- Gemma 4 31B
Gemma 4 31B Dense flagship VLM, 256K context, thinking mode.
- Gemma4-E2B
Gemma 4 E2B compact VLM, 128K context, bf16.
- Qwen3.6-35B-MoE
Qwen 3.6 35B/A3B MoE VLM, 256K context. fast inference.
- Qwen3.8-27B
Qwen 3.8 27B Dense VLM, 256K context, bf16.