Meet Clover-35B-A3B-Preview 🍀 the first open-weight production diffusion language model (dLLM) with 35B total parameters capable of 680 tps on a single card at BF16. Clover exhibits strong performance against both dLLM and AR models and is available today at @huggingface @ModelScope2022 and at our website

The combination of high TPS and intelligence makes Clover ideal for local llm agentic workloads. Each model call takes significantly less time to complete, making agents respond faster and with higher accuracy. Look out for our local deployment guide coming soon.

DiffuSpace is working with partners to explore cutting edge use cases such as drone logistics and navigation. If that sounds interesting, our DMs are open

Clover's combination of intelligence and speed makes it ideal for emerging use cases like generative UI, where applications can generate the user interface contextually and in realtime

Across six agent benchmarks Clover achieves the highest scores among the evaluated models: 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1, and 65.3 on τ³ Overall, outperforming fellow diffusion language models as well as similarly sized autoregressive models

Code for using Clover (600+ tps) for self-speculative decoding with SGLang and llama.cpp is available at github.com/DiffuSpace/Clover…

For more information about our model, see our blog at diffuspace.ai/blog/clover-35… and try it online at chat.diffuspace.ai/

In agentic workloads where time spend in every loop adds up, Clover's combination of intelligence and speed delivers significantly better user experience.

it's quiet... too quiet @ huggingface.co/Diffuspace

Excited to partner with @Acrab_Ai to bring diffusion language models to on-device agents. Up to 5× faster agent execution. From AI PCs to robotics and beyond. A step toward faster, more efficient on-device intelligence. ⚡

Intelligence should not only move forward token by token. It should think globally, evolve through uncertainty, and shape possibilities into answers. DiffuSpace is building diffusion language models as a foundation for the next generation of intelligence. Through parallel generation, bidirectional modeling, and iterative refinement, we aim to move beyond the constraints of autoregressive generation toward intelligence that is more flexible, efficient, and controllable. Shaping structure from possibilities, and exploring intelligence beyond the known.