@DiffuSpace_Labs introduces Clover-35B-A3B-Preview, an open-source diffusion language model with 35B total parameters, only about 3B active per token, and a 128K-token context window.
🤖 https://modelscope.ai/models/Diffuspace/Clover-35B-A3B-Preview

🏆 SOTA for open diffusion language models at release: On the benchmarks reported in the team’s launch blog, Clover leads the evaluated open-source diffusion language models.
⚡ Generate blocks, not just tokens: Block diffusion begins with masked token blocks and resolves high-confidence positions in parallel, rather than predicting strictly one token at a time.
🧠 Built on a sparse MoE backbone: Derived from Qwen3.6-35B-A3B, Clover keeps 35B total parameters while activating roughly 3B parameters per token.
📚 128K context for longer work: Supports up to 131,072 tokens across the prompt and response.
🛠️ Ready for experimental serving: Use the provided Transformers implementation for standalone PyTorch inference, or the official SGLang and llama.cpp backends for self-speculative decoding and serving.
📜 Apache-2.0 License. This is a preview release; the BF16 weights require approximately 69 GB of storage, plus inference memory.