Diffusion language models have spent most of their short history as small research systems or closed commercial products. DiffuSpace wants to change that. On October 30 the startup released Clover-35B-A3B-Preview, an open-weight mixture-of-experts model with 35 billion total parameters, of which about 3 billion are active for each token, and published the weights on Hugging Face and ModelScope with code on GitHub under the Apache-2.0 license.
The team has a track record in this niche. According to DiffuSpace, several of its members built Dream 7B at the University of Hong Kong's NLP group, a diffusion model released in April 2025 whose family has since passed 2.7 million downloads on Hugging Face. Clover is the group's first model under the new company, and its launch post frames the question directly: whether diffusion models stay competitive, and keep their speed advantage, once they are scaled up.
Clover is not trained from scratch. Its model card says it is a text-only derivative of Alibaba's Qwen3.6-35B-A3B and keeps that model's hybrid backbone of Gated DeltaNet and full-attention layers, with 40 layers, 256 routed experts, eight experts per token and one shared expert. DiffuSpace's changes target block diffusion, in which the model fills in a block of masked tokens in parallel and then treats the finished block as fixed context for the next one. Earlier text is compressed into the recurrent state of the DeltaNet layers, tokens inside the current block attend to each other in both directions, and the readout from that compressed memory competes in the same softmax as the block's own tokens. The company also describes a learned calibration for block size and a separate convolution branch for right-hand context. The context window is 128K tokens, and the model supports a thinking mode.
The headline numbers are DiffuSpace's own. For agentic coding it reports a SWE-bench Verified score of 73.6, a Terminal-Bench 2.1 result of 51.7 and a τ³ agent score of 65.3, which places it ahead of the diffusion models LLaDA2.2 flash and Inception's Mercury 2.5 as well as Qwen3.6-35B-A3B and Qwen3.5-35B-A3B in its comparison. Its general results are listed as 83.8 for GPQA Diamond, 95.8 for ZebraLogic and 72.0 for AA-LCR. For speed, the company says Clover averaged 680 tokens per second on internal agent workloads on a single NVIDIA GB300 at BF16 with a batch size of one, and 515 tokens per second on an H200 against 365 for a same-size Qwen setup accelerated with DFlash2.
Deployment options go beyond data-center GPUs. DiffuSpace says an INT4 build ran on an RTX 5070 Ti with 16GB of memory, reaching 101.8 tokens per second on MATH-500 compared with 48.5 in its tests of similar-size models. Official GGUF files for llama.cpp range from 18.24 GiB to 35.26 GiB and use the model's own diffusion output as a draft for self-speculative decoding, so no separate draft model is needed. The full BF16 weights need roughly 69 GB of GPU memory. The GitHub repository says its SGLang and llama.cpp backend code and setup guides are still to be added.
The release also previews what comes next. DiffuSpace describes a decision model built on Clover with a Joint Mode that answers up to eight related questions in one denoising pass, and claims it outscored TypeSafe's Jev on an internal set of linked legal-style judgments; that model is not yet released. The company says a harness for running agents and a coding app for desktop computers are in development, and it has named SigmaZ as an education partner running a closed beta.
Early interest is visible but modest: on October 30 the main model page showed about 1,500 downloads and 164 likes. None of the benchmark or speed figures has been checked independently, several rely on internal workloads, and the Preview label signals that DiffuSpace expects further changes. Still, Clover gives developers an openly licensed diffusion model at a scale that, until now, was mostly available only through closed APIs, which makes it a practical test of whether parallel decoding can pay off in agent loops.