chameleon-lizard/buckwheat-diffusion-base

Diffusion tool caller 150M — long pretrain

Code and reproduction: github.com/chameleon-lizard/diffusion-toolcaller

This is an inference-only checkpoint of a custom 149.7M-parameter masked diffusion language model after approximately 57.0B input tokens. The count includes the original ~3B pretrain tokens and two passes over a 27B-token corpus; it does not mean 57B unique tokens.

The continued-pretraining mixture targeted 45% FineWeb-Edu, 20% DCLM, 25% Stack-Edu Python/TypeScript/JavaScript, and 10% Cosmopedia-v2. Maximum context length is 2048. FP32 RoPE calculations and the original masked diffusion objective were retained.

This checkpoint has no tool router or slot heads. It is not an autoregressive chat model or a transformers.AutoModel package. Load it with hub.py in the code repository:

from hub import load_hub_model
model, _, config = load_hub_model("path/to/downloaded/model")
model.eval()

Files are model.safetensors, config.json, and the exact training tokenizer.json (SHA-256 6c8aaa9a542084f2457eab775d4eeb51f92a70c0fd9de28d5edb0ddec3c08d30). Optimizer and RNG state, absolute paths, and raw training data were removed. The tokenizer's upstream license has not been independently established here.

See the tool-calling posttrain for an evaluated checkpoint. This is a research release, not a tool-execution system.