{"id":22580,"date":"2010-03-29T16:20:58","date_gmt":"2010-03-29T16:20:58","guid":{"rendered":"https:\/\/scannn.com\/github-microsoft-mage-%c2%b7-github\/"},"modified":"2010-03-29T16:20:58","modified_gmt":"2010-03-29T16:20:58","slug":"github-microsoft-mage-%c2%b7-github","status":"publish","type":"post","link":"https:\/\/scannn.com\/lv\/github-microsoft-mage-%c2%b7-github\/","title":{"rendered":"GitHub - microsoft\/Mage \u00b7 GitHub"},"content":{"rendered":"\n
\n
\n<\/a>\n<\/div>\n\n

Mage<\/strong> is a family of lightweight, research-friendly multimodal models built at a fixed 4B-parameter<\/strong> budget. It is designed to make advanced visual understanding<\/strong> and generation<\/strong> accessible for controlled experiments, post-training research, and vertical-domain applications under realistic compute budgets.<\/p>\n

The family is organized around a shared codec-aligned efficiency<\/strong> philosophy \u2014 spend representation capacity where the signal is<\/em> \u2014 applied to both the understanding and the generation side:<\/p>\n

<\/p>\n\n\n\n\n\n\n
Model<\/th>\nTask<\/th>\nScale<\/th>\nCode<\/th>\nReport<\/th>\n<\/tr>\n<\/thead>\n
Mage-VL<\/a><\/strong><\/td>\nImage & video understanding, proactive streaming<\/td>\n4B<\/td>\nmage_vl\/<\/code><\/a><\/td>\nComing Soon<\/em><\/td>\n<\/tr>\n
Mage-Flow<\/a><\/strong><\/td>\nText-to-image generation & instruction-based editing<\/td>\n4B<\/td>\nmage_flow\/<\/code><\/a><\/td>\nPDF<\/a><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

<\/markdown-accessiblity-table><\/p>\n

Both models are compact enough to train, fine-tune, and deploy on modest hardware, yet remain competitive with much larger open systems in their respective domains.<\/p>\n\n

\n

Mage-VL \u2014 codec-native streaming vision\u2013language<\/h2>\n

<\/svg><\/a><\/div>\n

Mage-VL<\/strong> is a codec-native, proactive-streaming multimodal foundation model for image & video understanding, trained entirely from scratch<\/strong> at a compact 4B<\/strong> scale.<\/p>\n

Coming soon<\/strong> \u2014 code, checkpoints, and full details are on the way. Stay tuned.<\/p>\n

\u2192 mage_vl\/README.md<\/code><\/a><\/strong><\/p>\n

\n

Mage-Flow \u2014 efficient native-resolution generation & editing<\/h2>\n

<\/svg><\/a><\/div>\n

Mage-Flow<\/strong> is a compact 4B generative stack for text-to-image generation<\/strong> and instruction-based image editing<\/strong>, built from two co-designed components: Mage-VAE<\/strong> (a lightweight, high-fidelity latent tokenizer) and a Native-Resolution Multimodal Diffusion Transformer<\/strong> trained with rectified flow matching. Each task ships in Base<\/strong>, RL-aligned<\/strong>, and 4-step Turbo<\/strong> variants.<\/p>\n

Highlights<\/strong><\/p>\n