# Fara1.5-27B (slim)

> Fara1.5-27B is Microsoft's open-weight computer use model for web browsers, 27B parameters fine-tuned from Qwen3.5-27B. It reads screenshots and returns one click, keystroke or navigation at a time, pausing before irreversible steps. MIT, English only.

- Full: https://www.anchorterminal.com/tools/fara1-5-27b.md (~7,250 tokens) · this version ~1,730 tokens · JSON https://www.anchorterminal.com/tools/fara1-5-27b.json · canonical https://www.anchorterminal.com/tools/fara1-5-27b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/fara1-5-27b#pricing · reviews https://www.anchorterminal.com/tools/fara1-5-27b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #62 of 160 open-weight models · #4 in Desktop & mobile app control · not agent-ready · confidence low**


Assessment: A browser agent model with unusually careful documentation: what it does, what it's trained to stop at, what it gets wrong and how to box it in, under MIT. Treat it as a component for Microsoft's harness, in a sandbox, with a person on the irreversible steps. Safety testing is described but not published, and it's English only.

## Facts

- Kind: Open-weight model · vendor: Microsoft · category: Desktop & mobile app control · legal entity: Microsoft · provenance 67/100
- Packages: huggingface `microsoft/Fara1.5-27B`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/microsoft/Fara1.5-27B, read from the model card, the files and their metadata, the config, the chat template, the commit history, bartowski's GGUF repository and the runtimes' own code at their latest releases on 10 November 2026. Microsoft's company facts come from our other Microsoft listings. The paper, the fara repository beyond the presence of its security policy, MagenticLite and the Azure Foundry page weren't read. Nothing was downloaded or run
- Model: 27,356,728,560 parameters in the files. A multimodal decoder-only model from Microsoft Research AI Frontiers, fine-tuned with supervision from Qwen3.5-27B on trajectories from FaraGen1.5, Microsoft's multi-agent data pipeline. Qwen3_5ForConditionalGeneration
- Input and output: In: the user's goal, the latest screenshot or screenshots and the earlier thoughts and actions. Out: a chain-of-thought block, then one tool_call block holding a JSON object that calls computer_use with an action: clicks and drags at pixel coordinates, typing and keys, scrolling, visit_url, web_search, ask_user_question, wait or terminate. The card says it was mostly trained at a 1440×900 screen
- Context: 262,144 tokens, per the card's details table and config.json. The card's loop keeps only the most recent screenshots in the history
- Files: Safetensors in BF16 (54.71 GB), chat_template.jinja, config, processor and tokenizer files. No pickle files and no Python. Hugging Face lists a SHA-256 for the large files
- Licence: MIT, in the card's metadata and its licence section. There is no LICENSE file in the model repository. The base model, Qwen3.5-27B, is Apache 2.0
- Runtimes: The card documents vLLM. Released vLLM 0.33.0, SGLang 0.5.23 and Transformers 5.22.2 run the Qwen 3.5 architecture from the safetensors files. llama.cpp has the architecture too, and Microsoft publishes no GGUF, so the GGUF files on the Hub are bartowski's
- Recommended settings: The card's system prompt, word for word, temperature 0 and up to 2,048 output tokens in its example loop, with up to 10 images a prompt on vLLM
- Hardware: No memory figure on the card. The files are 54.71 GB in BF16. Trained on 64 NVIDIA B200 GPUs for 6 days
- Training data: Synthetic web-task trajectories kept only when a verifier agent judged them successful, grounding sets with coordinates, UI understanding data from page screenshots, and refusal data for harmful tasks. About 1 billion text tokens and under 1 billion images, English only, with the last data acquired on 20 March 2026, per the card

- Scores: Capability pending, Reach 55, Risk 75, Agent interface 80, Documentation and upkeep 65 · total over the 7 assessed categories
- Why:  · Reach, Memory 10, the only build from Microsoft is the BF16 safetensors, 54.71 GB, and 54.71 GB × 1.2 = 65.7 GB, which fits the reference machine b… · Risk, File safety 25, safetensors only, no .bin, .pt or .pth files and no Python in the repository. · Agent interface, Chat template 20, chat_template.jinja in the repository (7,756 characters), Qwen 3.5's template. · Documentation and upkeep, Model card 20, Intended use, out-of-scope use and known limitations, the architecture, the parameters, the context, the language (English on…
- Sources: 13, open questions: 6, both in the full twin
- Capabilities: browser.control, desktop.control, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/fara1-5-27b.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: https://github.com/microsoft/fara/tree/main/mage_vl_cowork; more in the full twin

## Before you call it

1. Run it inside Microsoft's harness (the fara CLI or MagenticLite), in a sandboxed browser with an allow-list. The card says not to give it open access on a machine with anything sensitive
2. Send the card's system prompt word for word. The model was trained against it
3. Match a 1440×900 screen for reliable clicks
4. Expect it to stop and ask before purchases, sign-ins, messages or personal data unless the user authorised them up front. Pass the question to the user; don't answer for them
5. Treat page text as untrusted. The model can be redirected by instructions on a page
6. It emits one action per turn at temperature 0, and ends with `terminate`

## Connect

```bash
pip install "vllm>=0.33.0"
vllm serve microsoft/Fara1.5-27B --dtype bfloat16 --max-model-len 262144 --limit-mm-per-prompt image=10
# then drive it with Microsoft's harness: github.com/microsoft/fara, or MagenticLite (github.com/microsoft/magentic-ui)
```

```bash
# vLLM serves an OpenAI-compatible API at http://localhost:8000/v1. Send the card's system prompt word for word,
# the goal and the latest screenshot as an image part, at temperature 0. The reply ends in one tool_call block
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Fara1.5-4B | C | 61 | browser.control, desktop.control, model.open-weights | https://www.anchorterminal.com/tools/fara1-5-4b.min.md |
| Fara1.5-9B | C | 61 | browser.control, desktop.control, model.open-weights | https://www.anchorterminal.com/tools/fara1-5-9b.min.md |
| Robocorp | B | 65.5 | desktop.control, browser.control | https://www.anchorterminal.com/tools/robocorp.min.md |
| Mage-VL-Cowork | C | 61 | desktop.control, model.open-weights | https://www.anchorterminal.com/tools/mage-vl-cowork.min.md |
| Browser Use | A | 80.3 | browser.control | https://www.anchorterminal.com/tools/browser-use.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)