Model • SGLang • Docs • Discord

LFM Diffusion

Block-diffusion language models from Liquid AI. LFM2.5-350M-Diffusion is LFM2.5-350M converted into a uniform-state block-diffusion model: it denoises 32-token blocks in parallel and decodes several times faster than its autoregressive parent at low batch size.

Note

lfm2.5-350m-diffusion-exp is an experimental release. Training, fine-tuning, TRL and native Transformers generate() support will be added to this repository.

Install

git clone https://github.com/Liquid4All/lfm-diffusion && cd lfm-diffusion
uv sync --extra cuda   # NVIDIA
uv sync --extra rocm   # AMD

Inference

PyTorch (reference implementation, any GPU):

uv run scripts/generate.py --prompt "Give three tips for getting better sleep." --nfe 8

from lfm_diffusion import generate, load_model

model, tokenizer = load_model("LiquidAI/lfm2.5-350m-diffusion-exp")
messages = [{"role": "user", "content": "Give three tips for getting better sleep."}]
print(generate(model, tokenizer, messages, config="nfe8", max_new_tokens=256))

SGLang (fast serving, OpenAI-compatible): see docs/serving.md.

Decoding

--nfe sets the denoising steps per 32-token block. Fewer steps decode faster at some cost in quality.

Preset	Steps per block	Use
nfe32	32	Highest quality
nfe8	8	Recommended balance
nfe4	4	Fastest

All presets use ancestral sampling with a temperature anneal; do not decode greedily. The YAML files in lfm_diffusion/decode_configs list every parameter.

Evaluation

See docs/evaluation.md to evaluate the model through the SGLang endpoint with lm-evaluation-harness and the official harnesses of the remaining benchmarks.

Citation

@inproceedings{tafreshi2026blockdiffusion,
  title     = {From Autoregression to Block Diffusion: Adapting Language Models for Efficient Parallel Decoding},
  author    = {Amin Tafreshi, Rouzbeh and Fan, Jack and Mosca, Edoardo and Lechner, Mathias and Amini, Alexander},
  booktitle = {NeurIPS 2026 Workshop},
  year      = {2026}
}

License

Code: Apache 2.0. Model weights: LFM Open License v1.0.