NewsPublished:
DiffuSpace Releases Clover-35B, an Open Diffusion Language Model Built for Speed
The Apache-licensed model writes text in parallel blocks instead of word by word, and a 4-bit build runs on a 16GB gaming GPU.
Almost every chatbot you have used writes the same way: one token after another, left to right. DiffuSpace, a team that grew out of the University of Hong Kong group behind the Dream 7B model, wants to change that. On October 30 it released Clover-35B-A3B-Preview, an open diffusion language model that drafts whole blocks of text at once and refines them in parallel, closer to how image generators work.
The numbers
- 35 billion parameters in a mixture-of-experts design, with only 3 billion active per step, plus an optional thinking mode
- Released under an Apache license, with weights on Hugging Face and ModelScope, code on GitHub, and support for SGLang and llama.cpp
- About 680 tokens per second on agent tasks on a single NVIDIA GB300, and 515 tokens per second on an H200
- An INT4 build that runs on an RTX 5070 Ti with 16GB of VRAM
DiffuSpace says Clover beats similarly sized step-by-step models such as Qwen3.6-35B-A3B on coding and tool-use tests, including 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1. Those are the company's own results, so independent testing will tell the real story.
Why it caught my eye
Speed matters most for agents, the setups where a model clicks around a browser or edits code in long loops. Every second saved per step adds up. The demos in the announcement include a model building a working mock desktop in a single web page and interactive 3D science explainers, which hints at quick interactive tools rather than just chat.
For hobbyists, the headline is that a model of this class fits on a gaming GPU. If diffusion text models keep closing the quality gap, local assistants for writing, translation notes or fan projects could get a lot snappier. The full write-up is on the DiffuSpace Hugging Face blog.
Source: huggingface.co