Microsoft's newly open-sourced 4B ultra-lightweight model, Mage-VL-Cowork, can actually run on a budget graphics card with just 4GB of VRAM to automatically write Python code for data organization, and even directly understand screen layouts to simulate mouse operations. In this issue, we also bring you Clover-35B-A3B-Preview, a diffusion language model that scored a whopping 73.6 on SWE-bench. How does it achieve an extreme speed of over 100 tokens per second? We will tell you in just a moment!
Featured This Week
1. microsoft/Mage-VL-Cowork: An Ultra-Lightweight Local Hybrid Computer-Use Agent
- What it is: This is an ultra-lightweight 4B (4 billion parameters) multimodal computer-use Agent model open-sourced by Microsoft under the developer-friendly apache-2.0 license. Its biggest highlight is its ability to autonomously switch between graphical user interface (GUI) and command-line interface (CLI) operations within a single decision-making mechanism.
- Use Cases:
- Scenario 1: Automating repetitive data organization tasks. When software engineers need to merge complex fields across multiple spreadsheets, they previously had to click through software manually. Now, Mage-VL-Cowork can directly write and execute Python scripts (such as calling the openpyxl library) in the terminal, efficiently handling file editing in the background via code and bypassing tedious GUI clicks.
- Scenario 2: Operating web or GUI applications without APIs. When developers need to control closed systems or legacy software, the model automatically switches to the computer_use pipeline, recognizing screenshots and simulating mouse clicks and keyboard inputs to complete interface operations just like a human.
- Scenario 3: Running system-level automation tasks in secure, isolated environments. If a team needs to execute system-level tasks without leaking sensitive code or trade secrets, they can fully deploy this model locally in a sandbox, Docker container, or virtual machine because of its compact 4B size. This allows the Agent to execute arbitrary shell commands and file operations in a secure, privacy-preserving environment.
- Hardware Requirements: The official model card lists its full BF16 weight size at around 9.48 GB. However, according to evaluations by the third-party analysis platform modelfit.io, its Q4_K_M quantized GGUF format combined with the vision projector has a total size of only about 3 GB. Therefore, an NVIDIA graphics card with at least 4 GB of VRAM (such as a GTX 1650 or RTX 3050) is sufficient to run it, and it will run perfectly fine on a 16GB Mac.
- Current Traction: It currently has a trending score of 104 on Hugging Face, with 752 cumulative downloads and 105 likes.
- Comparison with Peers: According to Microsoft's official data on the OSWorld-Verified benchmark, Mage-VL-Cowork achieved a 66.0% task success rate. This score not only surpasses the larger Muse Glimmer-30B (65.9%) and Qwen3.6-27B (63.9%), but also far outperforms Fara1.5-4B (13.9%), which is also a 4B model but solely focused on GUI.
- How to Get It: The official README states that developers can load the model directly via Transformers (requiring
trust_remote_code=Trueto be configured in the code) or deploy it using the official SGLang container. - Link: microsoft/Mage-VL-Cowork
2. Diffuspace/Clover-35B-A3B-Preview: The High-Speed Diffusion Powerhouse for the Strongest Local Coding Agent
- What it is: This is a 35B Mixture of Experts (MoE) diffusion language model released by DiffuSpace under the apache-2.0 license. It is currently the first open-source diffusion language model, deeply optimized for Agent tasks, and particularly excels at autonomous coding, tool calling, and web browsing.
- Use Cases:
- Scenario 1: Building a local autonomous Coding Agent. When software engineers want to build a local assistant that can automatically fix bugs in GitHub projects, Clover-35B-A3B-Preview demonstrates an extremely high task completion rate. It can autonomously write code, call external tools, and browse the web to efficiently resolve repository issues.
- Scenario 2: Constructing ultra-low latency local AI assistants. If developers need extremely fast response times to handle complex Agent workflows, the "Block Diffusion" technology used by this model can denoise and generate multiple tokens in parallel, providing a smooth generation experience with ultra-high throughput in complex automation pipelines.
- Hardware Requirements: According to the third-party deployment guide on mindstudio.ai and the model specifications, the full BF16 weights require about 69GB of VRAM. However, the official README mentions that they already provide 8-bit and 4-bit GGUF quantized versions, allowing developers to easily run local inference via llama.cpp. Additionally, the model supports Self-Speculative Decoding for further acceleration.
- Current Traction: It currently has a high trending score of 195 on Hugging Face, with 2.4k cumulative downloads and 199 likes.
- Comparison with Peers: According to official data from DiffuSpace, Clover-35B-A3B-Preview achieved an impressive score of 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1, outperforming the autoregressive model Qwen3.6-35B-A3B of a similar parameter scale. Furthermore, the official team notes that its generation speed for Agent tasks reaches 515 tokens/s on an H200, which is significantly faster than the Qwen model using DFlash2 speculative decoding (365 tokens/s).
- How to Get It: The official README states that developers can load it by installing llama.cpp and starting an OpenAI-compatible local server in the terminal:
llama serve -hf Diffuspace/Clover-35B-A3B-Preview-GGUF:Q4_K_M. - Link: Diffuspace/Clover-35B-A3B-Preview
Conclusion
From this week's selection, it is clear that the development of local Agent models has entered a brand-new phase of high efficiency and multi-modal operations. Whether it is Mage-VL-Cowork, which flexibly switches between GUI and CLI with just 4GB of VRAM, or Clover-35B-A3B-Preview, which supercharges generation speeds using block diffusion technology, both empower engineers to build highly productive local automated workflows while keeping their data completely private. That is all for today's introduction, see you in the next episode!
Comments