Fara1.5-27B-FP8

Fara1.5-27B-FP8

Fara1.5-27B-FP8 is a FP8 dynamic quantized version of microsoft/Fara1.5-27B, a multimodal computer use agent (CUA) for web browsers developed by Microsoft Research AI Frontiers. This quantization reduces model size and memory footprint while preserving the model's action-grounding and web-task-completion capabilities, making single-GPU deployment more accessible.

Fara1.5-27B is supervised fine-tuned from Qwen3.5-27B on synthetic web-task trajectories generated by FaraGen1.5. It observes the browser through screenshots and emits structured tool calls (click, type, scroll, visit URL, web search, and so on) to complete tasks end-to-end. It is vision-only at perception time, with coordinate-grounded actions and a critical-points safety design that pauses before personal info entry, payments, submissions, sign-ins, and other irreversible actions.

Quantization Details

Quantization was performed using llmcompressor with the following recipe:

default_stage:
  default_modifiers:
    QuantizationModifier:
      targets: [Linear]
      ignore: ['re:.*lm_head', 're:.*embed_tokens$', 're:.*visual.*', 're:.*model.visual.*',
        're:.*linear_attn.*']
      scheme: FP8_DYNAMIC
      bypass_divisibility_checks: false

Linear layers are quantized to FP8 with dynamic per-tensor activation scaling. The lm_head, embedding table, all vision-tower (visual) components, and linear_attn layers are excluded from quantization and remain at full precision to preserve grounding accuracy and multimodal fidelity.

Model Details

Use with vLLM

Fara1.5-27B-FP8 is served through vLLM with native support for compressed-tensors FP8 checkpoints.

Requirements

• torch >= 2.11.0
• vllm >= 0.19.1
• A GPU with FP8 support recommended (Hopper or Blackwell class) for best throughput; also runs on Ampere with FP8 dequantized on the fly.

Serve

vllm serve prithivMLmods/Fara1.5-27B-FP8 \
  --max-model-len 262144 \
  --limit-mm-per-prompt image=10

Client request

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

screenshot_b64 = capture_browser_screenshot()  # base64 PNG

messages = [
    {"role": "system", "content": FARA_SYSTEM_PROMPT},
    {"role": "user", "content": [
        {"type": "text", "text": "Book a table for 2 at a sushi place in Sunnyvale for Friday 7pm."},
        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{screenshot_b64}"}},
    ]},
]

response = client.chat.completions.create(
    model="prithivMLmods/Fara1.5-27B-FP8",
    messages=messages,
    temperature=0.0,
    max_tokens=2048,
)

print(response.choices[0].message.content)

The assistant reply contains chain-of-thought text followed by a <tool_call>{"name": "computer_use", "arguments": {...}}</tool_call> block. Execute the action in a sandboxed browser, capture the new screenshot, append the assistant turn and a tool/observation turn, and loop until the model emits action="terminate". Keep only the most recent 3 screenshots in the chat history. The screen resolution Fara is most commonly trained with is 1440x900; match this in your sandbox for the most reliable grounding.

System prompt

Use the original Fara system prompt verbatim for best results, as documented in the base model card.

Intended Use and Limitations

Intended use, known limitations, safety considerations, and responsible AI guidance are unchanged from the base model. See the Fara1.5-27B model card for full details on critical-points safety design, evaluation results, training data, out-of-scope use, and responsible AI considerations. Quantization to FP8 may introduce minor numerical differences relative to the bf16 base model; downstream accuracy should be validated for production use cases.

License

Released under the MIT License, consistent with the base model.

Papers

Fara-1.5: Scalable Learning Environments for Computer Use Agents — Hugging Face Papers