The story so far
Weights, every dated article, and repo commits on one rail.
8 repo commits since the weights landed
Teal is the weights (2026-07-25), grey ticks are repo commits.
Mage-VL
microsoft Apache 2.0 out 3 months ago
4.74B parameters, images and text in, text out, 256K tokens of context.
Verified 2026-10-25 Rank #605 in LLMs · score 74 method v1 Sources 3 Downloads 30d 8.5k · #517 of 619 tracked
Do people use it?8.5kdownloads, 30 days#517 of 619 we track · below the median+10 likes, downloads −8% this week440 likesHugging Face2026-10-24
Where it's used
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
No API provider serves it, so its downloads are people running it themselves: 14 quantized builds and 4 Spaces.
- As a service0no API provider listed
- On own hardware14quantized builds
- Built on11 fine-tunes · 0 adapters · 0 merges
- In apps4Hugging Face Spaces
Top Spaces running it: microsoft/mage-vl-demo ♥ 109 · lienthealien/mage-vl-demo ♥ 1
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 16 GB card at INT4
About 2.8 GB at INT4, 5.7 GB at FP8, 11 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
What people built on it
15 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
15 derivatives on the Hub, most of them quantized
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Reference
Everything else, folded: the full tables, every source.
Benchmarks
20 rows
Benchmarks 20 rows
measured third-party leaderboard rowauthor-reported number from the model card or eval files (20)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported AI2D · Understand images and video | 83.16 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported BLINK · Understand images and video | 65.11 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported CC-OCR · Extract text from documents | 32.25 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported CV-Bench | 87.79 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported CV-Bench-2D | 82.13 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported CV-Bench-3D | 94.75 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported ChartQA · Understand images and video | 84.88 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported ChartQA · Understand images and video | 32.57 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported DocVQA · Extract text from documents | 95.14 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported DocVQA · Extract text from documents | 87.46 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported EmbSpatial | 82.67 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported InfoVQA · Extract text from documents | 80.33 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported MMBench · Understand images and video | 84.02 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported MMBench · Understand images and video | 82.04 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported MME-Perception | 1709.54 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported MME-RealWorld | 66.52 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported MMStar · Understand images and video | 67.32 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported OCRBench · Extract text from documents | 81.8 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported SeedBench (All) | 76.78 | Mage-VL-4B | model author | 2026-10-07 |
| author-reported TextVQA · Understand images and video | 77.28 | Mage-VL-4B | model author | 2026-10-07 |
What the card says
1 cited sentence
Summary generated, cited
The architecture utilizes a codec-native, proactive-streaming approach to achieve up to 3.5× wall-clock inference speedup over uniform frame sampling.[1]
Generated 2026-10-07 15:13 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.
Use-case evidence
5 use cases
What it is good at 5 use cases
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
- Understand images and videoauthor-reportedAuthor reports AI2D w/ Mask 83.16Author reports ChartQA 84.88Author reports TextVQA-val 77.28+8 more
- Extract text from documentsauthor-reportedAuthor reports DocVQA-val 95.14Author reports InfoVQA-val 80.33Author reports OCRBench 81.80+1 more
- Chat and follow instructionsstructuraldecision-model estimate p=0.70tag: conversationalchat template present
- Run on a laptop or phonestructuraldecision-model estimate p=0.82fits: single-gpu
- Read long documentsstructuralconfig: 262,144 context
Facts
21 fields
Listed in LLMs
Facts
- Mage-VL has 4.74B parameters. (safetensors metadata, observed 2026-10-24)
- Context length is 262,144 tokens (256K). (config.json max_position_embeddings, observed 2026-10-24)
- License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
- Created 2026-07-25, last modified 2026-08-10. (HF API, observed 2026-10-24)
| Pipeline tag | image-text-to-text |
|---|---|
| Library | transformers |
| Architectures | MageVLForConditionalGeneration |
| Model type | mage_vl |
| Dtype | bfloat16 |
| Params by dtype | BF16: 4,741,793,792 |
| Size band | M |
| MoE | no |
| Context length | 262,144 (config.json max_position_embeddings) |
| Papers | arXiv:2607.24904 |
| License | Apache 2.0 (apache-2.0) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: not in registry · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed
vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Hardware fit
3 estimates
Hardware fit ESTIMATE
Weights-only memory for Mage-VL: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 11 GB · Single GPU (<24 GB) |
|---|---|
| FP8 (1 byte/param) | 5.7 GB · Single GPU (<24 GB) |
| INT4 (0.5 bytes/param) | 2.8 GB · Edge (<4 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
All variants
15 variants
Lineage
quantized: 14finetune: 1
Top variants (15 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| JohnTdi/Mage-VL-GGUF | quantized | GGUF | — | — | apache-2.0 | JohnTdi | 2 | 6.4k |
| mlx-community/Mage-VL-8bit | quantized | -8bit | 4.74B | M | apache-2.0 | mlx-community | 8 | 2.2k |
| mlx-community/Mage-VL-OptiQ-4bit | quantized | -4bit | 4.41B | M | apache-2.0 | mlx-community | 3 | 1.1k |
| ajh-code/Mage-VL-FP8-W8A8-W8A16 | quantized | -FP8 | 4.74B | M | apache-2.0 | ajh-code | 0 | 1.1k |
| litert-community/Mage-VL | quantized | — | — | — | apache-2.0 | litert-community | 3 | 869 |
| sahilchachra/Mage-VL-AWQ | quantized | AWQ | 4.74B | M | apache-2.0 | sahilchachra | 0 | 928 |
| SandLogicTechnologies/mage-vl-GGUF | quantized | GGUF | — | — | apache-2.0 | SandLogicTechnologies | 1 | 861 |
| sr29/Mage-VL-mlx-4bit | quantized | mlx | 4.74B | M | apache-2.0 | sr29 | 1 | 426 |
| ajh-code/Mage-VL-XPO3-NVFP4-W4A4 | quantized | NVFP4 | 3.15B | M | apache-2.0 | ajh-code | 0 | 456 |
| sahilchachra/mage-vl-mxfp4-mlx | quantized | mxfp4 | 4.74B | M | apache-2.0 | sahilchachra | 0 | 209 |
| RKNNAI/RK1828-VLM-Mage-VL | quantized | GGUF | — | — | apache-2.0 | RKNNAI | 0 | 206 |
| sahilchachra/mage-vl-mxfp8-mlx | quantized | fp8- | 4.74B | M | apache-2.0 | sahilchachra | 0 | 198 |
| momo321654/LLaVA-OneVision-2-4B-p16m2-Instruct | finetune | — | 4.74B | M | apache-2.0 | momo321654 | 0 | 194 |
| sahilchachra/Mage-VL-NVFP4A16 | quantized | — | 4.74B | M | apache-2.0 | sahilchachra | 0 | 189 |
| nativ-community/Mage-VL-OptiQ-4bit | quantized | -4bit | 4.41B | M | apache-2.0 | nativ-community | 0 | 47 |
Repository changelog
8 commits
Repository changelog 8 commits
| Date | Commit | SHA |
|---|---|---|
| 2026-08-10 | docs: update citation (yang2026mage, arXiv:2607.24904) | d88b153 |
| 2026-08-06 | Fix <|image_pad|> misalignment when images= and videos= coexist (#3) | b4e7d5a |
| 2026-07-29 | docs: add install-free Transformers usage section (#2) | 5c78cab |
| 2026-07-29 | Point Mage-VL arXiv links to arxiv.org/abs/2607.24904 | ac32ce8 |
| 2026-07-27 | Restore data-efficient tokenizer finding (560M images) | 8484f31 |
| 2026-07-27 | Remove data-efficiency claims from model card | 64e9df0 |
| 2026-07-26 | Upload Mage-VL: unified codec-native streaming VLM (image+video understanding + proactive gate) | 12acbba |
| 2026-07-25 | initial commit | f3afe21 |
Papers
1 paper
Papers 1
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
Similar models
3 models
Similar models
- Qwen3.5-4B · 4.66B same size band (M), same task (image-text-to-text), more providers, permissive license
- gemma-3-4b-it · 4.30B same size band (M), same task (image-text-to-text), more providers
- Qwen3-VL-4B-Instruct · 4.44B same size band (M), same task (image-text-to-text), more providers, permissive license
Sources
3 sources
Sources
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:42 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:13 UTC · Last verified 2026-10-25
Follow Mage-VL
Watch this model
- /feeds/models/microsoft--Mage-VL.xml(RSS 2.0) · every change to Mage-VL
- /feeds/families/microsoft--Mage.xml(RSS 2.0) · every change across the Mage family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.