Fara1.5-9B PrismaAURA 5-bit
Mixed-precision PrismaAURA export of microsoft/Fara1.5-9B for NVIDIA Blackwell and vLLM compressed-tensors serving.
Official MagenticLite model pairing
Microsoft Magentic-UI is the officially recommended host application for this two-model stack. Microsoft describes MagenticLite as pairing an on-device-friendly orchestrator model, MagenticBrain, with a specialized browser-use model, Fara.
Use these two PrismaAURA repositories together for that pairing:
- Orchestrator model: MagenticBrain PrismaAURA 5-bit
- Browser-use model: Fara1.5-9B PrismaAURA 5-bit (this repository)
- Official host application: microsoft/magentic-ui
These are independently produced quantized derivatives of the Microsoft models; Magentic-UI, MagenticBrain, and Fara are the official Microsoft components and model roles.
Quantization
- AURA format menu: NVFP4 + FP8 E4M3 + BF16
- Selected language-body operating point: 5.000676 bits per quantizable parameter
- Assignment: 182 NVFP4, 58 FP8 E4M3, and 118 BF16 Linears
- Visual encoder: BF16 passthrough
- Held-out real KL vs BF16 during frontier selection: 0.056930117
- Source model: 9,409,813,744 parameters, BF16
- Exported weights: 8.67 GiB across two safetensors shards
The selected point was measured from the 4.75, 5.0, and 5.25 bpp candidates. Their surrogate/real-KL rank correlation was 1.0. Multimodal Fisher calibration used eight synthetic image-text samples; text calibration used 32 samples of 1,024 tokens.
Serve
vllm serve /path/to/model \
--quantization compressed-tensors \
--trust-remote-code \
--max-model-len 32768 \
--limit-mm-per-prompt image=10
The artifact was load-tested with vLLM 0.24.0 on an NVIDIA GB10. vLLM selected FlashInferCutlassNvFp4LinearKernel for NVFP4 and CutlassFP8ScaledMMLinearKernel for FP8, then completed a generation smoke test.
Fara is a computer-use agent. Run it only in a sandboxed browser environment with appropriate allow-lists, monitoring, and confirmation gates for irreversible actions.
- Downloads last month
- 10