On this page
DiffuSpace’s blog post dated October 28, 2026 announces Clover-35B-A3B-Preview—the first Clover release. It is an open-weight Mixture-of-Experts (MoE) diffusion language model with about 35 billion total parameters and about 3 billion activated parameters, thinking mode, and a 128K context window (up to 131,072 tokens on the Hugging Face card). The release is under Apache-2.0. Weights and related code are on Hugging Face, ModelScope, and GitHub, with inference support for SGLang and llama.cpp. Related open-weights reading: Atria Dawn Formal Weights Are Out: 512K Context on Hugging Face.
Lineage and positioning (vendor description)
DiffuSpace frames Clover as a follow-on to the HKU NLP Dream line (Dream 7B, April 2025). The Hugging Face model card lists base_model: Qwen/Qwen3.6-35B-A3B. The blog describes Clover as a MoE diffusion LLM aimed at keeping global-generation speed while scaling parameters, and points to comparisons against the same-size autoregressive Qwen3.6-35B-A3B—those head-to-heads are DiffuSpace’s own reported figures. The same post also sketches a block-diffusion redesign of Gated DeltaNet for parallel denoising—architectural detail is on the blog.
Agent scores and throughput (vendor-reported)
On agentic coding and tool-use tables in the blog, DiffuSpace reports SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7, and τ³ Overall 65.3. For inference speed, it cites about 680 tokens/s on a single NVIDIA GB300 GPU in BF16, batch size 1, on internal agent workloads (with separate H200 figures on the same post). All of those numbers are vendor-reported—for reference only. Related open-weights coding reading: Recursive Releases Flint, an Open-Weights Coding Model Built to Run on a Laptop.
Where weights are today—and what is not out yet
Weights live at Diffuspace/Clover-35B-A3B-Preview on Hugging Face and ModelScope; the GitHub repo is DiffuSpace/Clover-35B-A3B-Preview. DiffuSpace also points readers to chat.diffuspace.ai and an apply flow on diffuspace.ai for try-out access—framed as application / early access, not an ungated public API.
Separately, the blog introduces a decision model built on Clover-35B-A3B-Preview (independent mode for one question at a time; joint mode for up to eight related answers in one denoising pass). DiffuSpace states that decision model will be released at a later date—it is not part of today’s downloadable weights.
OpenRouter: the public models list still showed 464 entries and no clover / diffuspace model IDs. Today’s entry points are DiffuSpace’s listed Hugging Face / ModelScope / GitHub paths. Same-day open-weights multimodal reading: Liquid AI Releases LFM3: Open-Weight Multimodal Family Around 2B.
Availability, serving recipes, and apply flows follow DiffuSpace’s Clover blog, the Hugging Face card, and the GitHub README. For reference only.