Microsoft
GELab-Zero-4B-preview-Sico-Evolution is a open-weight Microsoft specialized model.
Running this yourself: consumer gpu should be enough.
---
Quality Score
---
Arena ELO
4B
Parameters
---
Context
Sign in to join the discussion
268
Downloads
41
Likes
Jun 2026
Released
Launches
2
API
1
Research
2
General
5
Recent launch, pricing, benchmark, and API signals linked to this model or its provider.
Thrilled to announce by popular demand MAI-Code-1-Flash is now generally available for GitHub Copilot Business and GitHub Copilot Enterprise - fast, efficient, and custom designed to help you build more for less on @github
View sourceLet’s go a bit deeper into Frontier Tuning launched at Build and see a live demo! Frontier Tuning is how we enable you to develop custom AI by building a reinforcement learning environment (RLE) to hill climb on your data, tools, and knowledge. Think of it like a training gym https://t.co/XhZHTuks39
View sourceWhat does it take to build coding models that meet developers where they work? Go behind the scenes with Microsoft AI to explore how we build and optimize code. From training and evaluation to performance, safety, and real-world developer feedback. https://t.co/kkzNkci2ru
View sourceVision-Language-Action (VLA) models can generalize across diverse manipulation tasks, but their imitation-learning-based policies remain brittle in precise physical interactions due to compounding execution errors; Can a reinforcement learning policy trained purely in simulation improve the robustness of real-world VLAs zero-shot? Residual RL, which learns a corrective policy on top of a frozen VLA, offers a natural framework, but existing approaches face a fundamental sim-to-real dilemma: privileged-state methods require lossy distillation for deployment; image-based methods suffer from the visual domain gap; and real-world RL is costly and unsafe. We propose an object-centric residual RL framework that refines VLA actions using object poses, enabling a compact observation space that transfers consistently between simulation and reality. To align the two domains, we additionally replay the same teleoperation demonstrations in simulation to train a sim counterpart of the real-world VLA. The residual RL policy is trained only in simulation with pose noise injection and dropout, and transfers zero-shot to the real robot. Across five manipulation tasks on a real Franka Research 3 (FR3) robot, our method improves the success rate from 42% to 76% zero-shot, and the improved rollouts can be further reused to retrain the base VLA for self-improvement without additional teleoperation. Project page: https://www.microsoft.com/en-us/research/articles/object-centric-residual-rl/
View sourceLarge Language Model (LLM) coding agents have achieved strong results on software engineering tasks, yet repository exploration remains a major bottleneck: locating relevant code consumes substantial token budget and pollutes the agent's context with irrelevant snippets. In most agents, the same model explores the repository and solves the task, leaving exploratory reads and searches in the solver's history. We present FastContext, a dedicated exploration subagent that separates repository exploration from solving. Invoked on demand, FastContext issues parallel tool calls and returns concise file paths and line ranges as focused context. FastContext is powered by specialized exploration models spanning 4B--30B parameters. We bootstrap them from strong reference-model trajectories and refine them with task-grounded rewards for broad first-turn search, multi-turn evidence gathering, and precise citation generation. Across SWE-bench Multilingual, SWE-bench Pro, and SWE-QA, integrating FastContext into Mini-SWE-Agent improves end-to-end resolution rates up to 5.5\% while reducing coding-agent token consumption up to 60\%, with marginal overhead. These results show that repository exploration can be separated from solving and handled effectively by specialized models. Code and data: https://github.com/microsoft/fastcontext
View sourceThrilled to announce by popular demand MAI-Code-1-Flash is now generally available for GitHub Copilot Business and GitHub Copilot Enterprise - fast, efficient, and custom designed to help you build more for less on @github
We shipped a new coding model built for your everyday dev work. MAI-Code-1-Flash is fast, token-efficient, and trained inside real GitHub Copilot environments. It plans, builds, runs, and tests. All from Copilot Chat in VS Code. Watch it go from a single frost banner to a full
MAI-Image-2.5 ranked #2 for text-to-image and #3 for image editing on @ArtificialAnlys - showing strong performance across both generation and precise image edits. From rainy-window blur to a clear, usable street scene, while preserving object consistency, lighting, https://t.co/q8MFDrr8uB
What happens when speech, transcription, and coding models work together? This prototype demo, built using a VS Code fork, showcases how MAI-Transcribe, MAI-Voice, and MAI-Code-1-Flash can work together in a unified workflow to transform spoken instructions into working code. https://t.co/z35l4fV2xQ
Behind every model is a team dedicated to solving difficult challenges, exploring new ideas, and continuously pushing technology forward. Meet some of the people behind Microsoft AI. Watch the full video here: https://t.co/oNxtle7FtO https://t.co/8fsVY6DnFQ
What does it take to build coding models that meet developers where they work? Go behind the scenes with Microsoft AI to explore how we build and optimize code. From training and evaluation to performance, safety, and real-world developer feedback. https://t.co/kkzNkci2ru
Let’s go a bit deeper into Frontier Tuning launched at Build and see a live demo! Frontier Tuning is how we enable you to develop custom AI by building a reinforcement learning environment (RLE) to hill climb on your data, tools, and knowledge. Think of it like a training gym https://t.co/XhZHTuks39
Fun fact – the images announcing MAI-Image-2.5 - our image generation and editing model were generated by the model itself! Strong photo-realism, precise edits, quality down to the fine details. Try it on our Playground: https://t.co/FyXSioHZKA or on @OpenRouter and share https://t.co/K7XgAmCBcs
Vision-Language-Action (VLA) models can generalize across diverse manipulation tasks, but their imitation-learning-based policies remain brittle in precise physical interactions due to compounding execution errors; Can a reinforcement learning policy trained purely in simulation improve the robustness of real-world VLAs zero-shot? Residual RL, which learns a corrective policy on top of a frozen VLA, offers a natural framework, but existing approaches face a fundamental sim-to-real dilemma: privileged-state methods require lossy distillation for deployment; image-based methods suffer from the visual domain gap; and real-world RL is costly and unsafe. We propose an object-centric residual RL framework that refines VLA actions using object poses, enabling a compact observation space that transfers consistently between simulation and reality. To align the two domains, we additionally replay the same teleoperation demonstrations in simulation to train a sim counterpart of the real-world VLA. The residual RL policy is trained only in simulation with pose noise injection and dropout, and transfers zero-shot to the real robot. Across five manipulation tasks on a real Franka Research 3 (FR3) robot, our method improves the success rate from 42% to 76% zero-shot, and the improved rollouts can be further reused to retrain the base VLA for self-improvement without additional teleoperation. Project page: https://www.microsoft.com/en-us/research/articles/object-centric-residual-rl/
Large Language Model (LLM) coding agents have achieved strong results on software engineering tasks, yet repository exploration remains a major bottleneck: locating relevant code consumes substantial token budget and pollutes the agent's context with irrelevant snippets. In most agents, the same model explores the repository and solves the task, leaving exploratory reads and searches in the solver's history. We present FastContext, a dedicated exploration subagent that separates repository exploration from solving. Invoked on demand, FastContext issues parallel tool calls and returns concise file paths and line ranges as focused context. FastContext is powered by specialized exploration models spanning 4B--30B parameters. We bootstrap them from strong reference-model trajectories and refine them with task-grounded rewards for broad first-turn search, multi-turn evidence gathering, and precise citation generation. Across SWE-bench Multilingual, SWE-bench Pro, and SWE-QA, integrating FastContext into Mini-SWE-Agent improves end-to-end resolution rates up to 5.5\% while reducing coding-agent token consumption up to 60\%, with marginal overhead. These results show that repository exploration can be separated from solving and handled effectively by specialized models. Code and data: https://github.com/microsoft/fastcontext