Microsoft Releases Mage-Flow, a 4B Open-Weight Model That Matches 20B+ Rivals on Image Generation and Editing
Microsoft has released Mage-Flow, a 4B-parameter open-weight foundation model for text-to-image generation and instruction-based editing. The company claims it matches or beats much larger open systems like Qwen-Image (20B) and FLUX.2 (32B) while running faster and using less memory.
Microsoft has released Mage-Flow, a 4-billion-parameter open-weight foundation model for text-to-image generation and instruction-based image editing, available now on Hugging Face. According to Microsoft, the compact model matches or exceeds the benchmark performance of open systems five to eight times its size, including Qwen-Image (20B), FLUX.2-dev (32B), and FireRed-Image-Edit (20B).
What was released
The Mage-Flow stack consists of two shared components: Mage-VAE, a one-step diffusion latent tokenizer, and NR-MMDiT, a 4B Native-Resolution Multimodal Diffusion Transformer trained with rectified flow matching. Six checkpoints are available, split across two model lines — Mage-Flow (text-to-image) and Mage-Flow-Edit (instruction-based editing) — each shipping in Base, RL-aligned, and 4-step Turbo variants.
| Model | Task | Steps | Hugging Face |
|---|---|---|---|
| Mage-Flow-Base | text→image | 30 | microsoft/Mage-Flow-Base |
| Mage-Flow | text→image | 20 | microsoft/Mage-Flow |
| Mage-Flow-Turbo | text→image | 4 | microsoft/Mage-Flow-Turbo |
| Mage-Flow-Edit-Base | editing | 30 | microsoft/Mage-Flow-Edit-Base |
| Mage-Flow-Edit | editing | 30 | microsoft/Mage-Flow-Edit |
| Mage-Flow-Edit-Turbo | editing | 4 | microsoft/Mage-Flow-Edit-Turbo |
Benchmark claims
On GenEval, Mage-Flow scores 0.90, ahead of Qwen-Image (0.87), Z-Image-Base (0.84), and FLUX.2-dev (0.87), all of which use far more parameters. On DPG-Bench, Mage-Flow posts 86.49 versus Qwen-Image's 88.32. For editing, Mage-Flow-Edit-Turbo scores 8.271 on GEdit-EN and 8.264 on GEdit-CN, ahead of FireRed-Image-Edit-1.0 (7.943/7.887) and Qwen-Image-Edit-2511 (7.877/7.819), according to Microsoft's reported figures. These numbers come from Microsoft's own technical report; they have not been independently verified.
Efficiency claims
Microsoft says Mage-VAE matches FLUX.2-VAE reconstruction fidelity while using roughly 12x fewer encode MACs and 22x fewer decode MACs per pixel. The company also reports that native-resolution packing combined with fused CUDA kernels cut per-step training time from approximately 1.93 seconds to 0.78 seconds — a claimed 2.5x training speedup. At 1024x1024 resolution on a single A100 GPU, Microsoft claims Mage-Flow-Turbo generates an image in 0.59 seconds and Mage-Flow-Edit-Turbo completes an edit in 1.02 seconds, with peak memory usage of 18-20 GB — described as the lowest among compared systems.
The model supports native resolution generation from 512 to 2048 pixels across arbitrary aspect ratios, including extreme 4:1 ratios such as 512x2048. Mage-Flow-Edit supports semantic content editing, appearance transformation, image restoration, and structure-aware editing within a single unified model, and can reportedly generate diverse outputs from a single reference image.
Pricing is not applicable — the model weights are freely available on Hugging Face under Microsoft's repository. Training data cutoff and license terms were not specified in the released materials.
What this means
Mage-Flow's core pitch is parameter efficiency: matching or beating models 5-8x larger through tokenizer-backbone co-design rather than brute-force scaling. If Microsoft's benchmark numbers hold up under independent testing, a 4B open-weight model that runs on a single A100 with sub-second inference and under 20GB of memory would meaningfully lower the barrier for developers who want competitive image generation and editing without renting multi-GPU clusters. The real test is whether third-party evaluations on GenEval, GEdit-Bench, and similar suites confirm Microsoft's self-reported scores, and whether the efficiency gains translate to production deployments rather than just controlled A100 benchmarks. For now, treat the head-to-head comparisons against Qwen-Image, FLUX.2, and FireRed-Image-Edit as vendor-reported until reproduced elsewhere.
Related Articles
Cloudflare releases Clairic, a single-pass document extraction model, in preview on Workers AI
Cloudflare has released Clairic, a model that extracts named fields from unstructured text in a single pass instead of generating tokens one at a time. It is in preview on Workers AI, with open weights promised soon. Cloudflare claims a multi-step agent chain using Clairic runs in 19.3 seconds, versus 61.7 seconds for GLM-5.3 alone.
MiniMax M3.1 Flash scores 42 on AA Intelligence Index at $0.15/$0.60 per 1M tokens, open weights Oct 23
MiniMax has released M3.1 Flash, a reasoning model that scores 42 on the Artificial Analysis Intelligence Index, up from 29 for M3. It costs $0.15 per 1M input tokens and $0.60 per 1M output tokens, and open weights are planned for October 23.
Mistral releases Timestral-209M, a time series model claiming top zero-shot scores on three benchmarks
Mistral AI has released Timestral-209M, the largest of four time series foundation models (5M, 28M, 123M, 209M parameters). According to the model card, it leads GIFT-Eval, TIME and fev-bench in zero-shot forecasting, with a context of up to 32,768 time steps and 21-quantile probabilistic output.
Taste-Bench: Best AI model scores 59.7% on long-term agent decisions, where random guessing yields 25%
Researchers from Microsoft and City University of Hong Kong released Taste-Bench, a 502-question benchmark that tests whether models can pick the better path at decision forks in long agent runs. The top model, GPT-5.6 Sol, scored 59.7% against a 25% guessing baseline. The authors also show a distilled Qwen3.6-27B student can lift SWE-bench Pro success from 14.6% to 33.7%.
Comments
Loading...