generated: '2026-08-16'
method: searched
source: >-
  https://github.com/furiosa-ai/furiosa-sdk/blob/main/python/furiosa-server/README.md,
  https://developer.furiosa.ai/latest/en/get_started/prerequisites.html,
  https://developer.furiosa.ai/latest/en/furiosa_llm/furiosa-llm-serve.html
summary: >-
  There is no hosted sandbox, no test-mode key and no trial console - there could not be, since
  FuriosaAI operates no API. What it ships instead is the hardware-vendor equivalent: an NPU
  SIMULATOR package that lets the stack run end to end without an RNGD card, plus small sample
  models and inputs committed to the SDK repo. That is the honest analogue of test cards here.
hosted_sandbox: false
test_mode_keys: false
simulator:
  package: furiosa-libnpu-sim
  distribution: FuriosaAI APT / RPM repository
  install: sudo apt install furiosa-libnpu-sim
  hardware_alternative: furiosa-libnpu-xrt
  evidence: >-
    The furiosa-server README's development setup reads
    "sudo apt install furiosa-libnpu-sim # or furiosa-libnpu-xrt if you have Furiosa H/W" -
    i.e. the simulator is the documented path for developing without a device.
  scope: >-
    Gen-1 (Warboy) era package documented in the furiosa-sdk repo. The current RNGD developer
    center documents no simulator equivalent for furiosa-llm; the 2026.x getting-started path
    assumes an installed RNGD device verified with `lspci -nn | grep FuriosaAI`.
sample_fixtures:
  - path: python/furiosa-server/samples/data/MNISTnet_uint8_quant.tflite
    description: Quantized MNIST TFLite model for smoke-testing the model server.
  - path: python/furiosa-server/samples/data/MNIST_inception_v3_quant.tflite
    description: Quantized Inception v3 model.
  - path: python/furiosa-server/samples/data/SSD512_MOBILENET_V2_BDD_int_without_reshape.tflite
    description: Quantized SSD512 MobileNetV2 detection model.
  - path: python/furiosa-server/samples/mnist_input_sample_01.json
    description: A ready-made KServe v2 inference_request body for the MNIST model.
sample_fixtures_repo: https://github.com/furiosa-ai/furiosa-sdk
smallest_reproducible_model:
  model: furiosa-ai/Qwen2.5-0.5B-Instruct
  source: >-
    Used as the model argument in the official container quick-start
    (`docker run ... furiosaai/furiosa-llm:latest serve furiosa-ai/Qwen2.5-0.5B-Instruct`).
    The smallest published artifact for verifying a serving install.
  requires: A Hugging Face access token (HF_TOKEN) and an RNGD device (--device /dev/rngd).
placeholder_values:
  - value: EMPTY
    where: the `model` field and OPENAI_API_KEY in every serving example
    note: >-
      Published verbatim by FuriosaAI, and meaningful rather than decorative - the server ignores
      the model value (one model per process) and runs without an API key by default.
test_clocks: false
webhook_replay: false
cookbook: https://github.com/furiosa-ai/furiosa-apps