Model releases · 11 August 2026
A browser agent that looks at screenshots and clicks, released at three sizes under one of the most permissive licences Microsoft has used — and built on a base that is not Microsoft's.
| Released by | Microsoft Research AI Frontiers |
|---|---|
| Published | July 2026 |
| Sizes | 4B, 9B, 27B |
| Base model | Qwen3.5 (Qwen3_5ForConditionalGeneration) |
| Purpose | Multimodal computer-use agent for web browsers |
| Layers / hidden size | 32 / 2560 (4B variant) |
| Context length | 262,144 tokens |
| Vocabulary | 248,320 tokens |
| Licence | MIT |
| Weight size | 9.11 GB for the 4B variant |
| Hosted at | microsoft on Hugging Face |
Fara 1.5 is a computer-use agent: it observes a web browser through screenshots and acts by emitting structured tool calls — click, type, scroll, visit URL, web search — to complete a task end to end. That is a narrower and more concrete job than general assistance, and it is the reason the models are multimodal.
Microsoft positions the 4B as viable on a single accelerator with room for screenshot history, which matters for an agent that has to remember what the page looked like several steps ago.
The configuration is explicit: Qwen3_5ForConditionalGeneration, with a 248,320-token vocabulary and a 262,144-token context — the Qwen3.5 signature. Microsoft Research built its computer-use agent on Alibaba's open base rather than on Phi, its own small-model line.
That is a reasonable read on where the Phi line stands: the most recent Phi releases predate this, and Microsoft's newer open work is appearing under the Fara and Mage names.
MIT across all three sizes, with no acceptable-use appendix. For a computer-use agent — a category where vendors often attach usage restrictions — that is notable.
| Runtime | Supported | Notes |
|---|---|---|
| transformers | Yes | Reference implementation from Microsoft. |
| llama.cpp | Partial | The underlying Qwen3.5 architecture is supported, but the screenshot-driven agent loop is not something a plain text runtime provides. |
| MLX | Partial | Community conversions only. |
| Vision input | Yes | Central to the design; the model reads browser screenshots. |
A family of multimodal computer-use agent models from Microsoft Research that drive a web browser by reading screenshots and emitting structured tool calls. Released at 4B, 9B and 27B.
MIT, across all three sizes, with no additional acceptable-use policy.
Qwen3.5. The configuration declares the Qwen3.5 architecture, vocabulary and context length rather than a Microsoft base such as Phi.
9.11 GB in the published safetensors form.
The underlying architecture is supported, but Fara is an agent that consumes screenshots and emits tool calls — a plain text inference runtime does not provide that loop.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 11 August 2026