tensorx-datasheet

An always-up-to-date Bifrost model datasheet that includes TensorX pricing and model information, refreshed daily and available as a raw file directly from GitHub:

https://raw.githubusercontent.com/agile-lab-dev/tensorx-datasheet/refs/heads/main/data.json

Point your Bifrost data sheet URL there and TensorX-hosted models show up with correct costs, token limits, and capabilities — no maintenance on your side.

Sample entry

{
  "deepseek/deepseek-v3.2": {
    "mode": "chat",
    "provider": "deepseek",
    "base_model": "deepseek-v3.2",
    "max_input_tokens": 163840,
    "max_output_tokens": 163840,
    "max_tokens": 163840,
    "input_cost_per_token": 3e-07,
    "output_cost_per_token": 5e-07,
    "cache_read_input_token_cost": 7.5e-08,
    "supports_function_calling": true,
    "supports_tool_choice": true,
    "supports_reasoning": true,
    "metadata": { "notes": "..." }
  }
}

How it works

A single Python script (tensorx-datasheet.py) fetches the live TensorX models API, rewrites each entry into the Bifrost datasheet schema, and concats it on top of the official Bifrost datasheet: all Bifrost models are kept, and for models present in both, the TensorX entry wins (fields are merged per-model, with TensorX values taking precedence). A daily job regenerates and commits data.json. To run it manually (requires uv):

uv run tensorx-datasheet.py --output data.json

TensorX model ids are namespaced with a tensorx/ prefix by default (e.g. tensorx/deepseek-v3.2). Use --prefix to change it, or pass an empty string to keep the raw model ids:

uv run tensorx-datasheet.py --output data.json --prefix mylab
uv run tensorx-datasheet.py --output data.json --prefix ""