Microsoft

A fine-tuned version of Qwen to explore and find the correct files for bigger coding models, which improves token efficiency while maintaining accuracy.

A Phi model which uses the SigLIP-2 vision encoder.

Another Computer-Use-Agent model from Microsoft. They've been releasing these models for a while.

An open version of NotebookLM, i.e., a model to generate multi-speaker "podcasts". However, Microsoft has removed the code and the largest model due to misuse. There are, however re-uploads of the model and forks of the code released under the same MIT license as the original models and code.

A version of the multimodal Phi for computer use. Microsoft has been working on CUA (Computer Use Agents) for a while in the open.

A hybrid model to speed up inference.

Microsoft has been working on GUI-grounded models for some time. This is their latest addition, which uses an attention-based action head to find the relevant object instead of outputting coordinates directly.

Phi also joined the ranks of the reasoning models. The "plus" variant uses GRPO on top of an SFT model. We found the model surprisingly capable in our testing considering both the size and the reputation of the Phi series as benchmark-maxers.

An update to Phi, supporting text, images and audio as input.

A VLM trained for various agent-based tasks, such as GUI navigation, robotics or gaming.

A model to generate bounding boxes for UI screenshots. Not only can it detect the position of icons, buttons, texts etc., but it can also parse the content and interpret whether the element is interactive, which is crucial for GUI agents.

Microsoft has released a "world model" to play a game, similar to oasis from the last episode. The game they used is the MOBA "Bleeding Edge", released in 2020. While this is meant as a tech demo, Microsoft intends to use this technology to preserve older games on new platforms. AI-based gaming might come sooner than expected.

Another entry in the Phi series, which cannot get rid of its reputation of performing well on benchmarks, while falling short in real-world usage.

A new MoE model series from Microsoft. It has some architecture tweaks, such as SparseMixer-v2, but it isn't clear what the differentiator is (like Phi).

Largely incremental updates to the Phi family, but the MoE addition is a good option.

Largely incremental updates to the Phi family, but the MoE addition is a good option.

Another member of the Phi family.

Reminder that Phi had a vision version!

We knew these models were coming, but they're solid for trying tasks like data filtering, local fine-tuning, and more on.

We knew these models were coming, but they're solid for trying tasks like data filtering, local fine-tuning, and more on.