DiffuSpace releases Clover-35B, an open diffusion language model preview

DiffuSpace has released Clover-35B-A3B-Preview, which it calls the first open-weight production diffusion language model, a type of model that generates text by denoising blocks in parallel rather than writing one token at a time the way standard autoregressive models do. The model has 35 billion total parameters with about 3 billion active per token, a 128K-token context window, and is derived from Qwen3.6-35B-A3B under an Apache 2.0 license.

The team says Clover reaches 680 tokens per second on a single card at BF16 precision, and that it performs strongly against both other diffusion models and standard autoregressive models on the benchmarks in its launch blog, though this is DiffuSpace's own reporting as a preview release.

Architecturally, Clover keeps the hybrid Gated DeltaNet and full-attention mixture-of-experts backbone from its Qwen base, with 40 layers, 256 routed experts, top-8 token routing, and one shared expert per layer, plus added learned parameters that let tokens attend within a block while preserving a causal prefix. Generation uses block diffusion with a technique called LowConfidence decoding, filling in the most confident token positions first within each block of 32 tokens by default. The model is available now on Hugging Face and ModelScope, along with code for self-speculative decoding through SGLang and llama.cpp.

Why it matters

A fast, open diffusion language model gives developers a real alternative generation architecture to try against standard autoregressive LLMs for latency-sensitive applications.

Source: DiffuSpace on X