Today for AI
HOT RADAR
llmHEAT 16.8°

AI CURATED HOT STORY · 10/28/2026

DiffuSpace Open-Sources Clover-35B, First MoE Diffusion LLM for Agents

2 reports archived2 independent sourcesupdated 10/29/2026, 10:18:28
Synthesis & Latest Updates
2 Sources Cross-Validated

Shenzhen-based DiffuSpace has open-sourced Clover-35B-A3B-Preview, the first large-scale diffusion language model (dLLM) to integrate a Mixture-of-Experts (MoE) architecture. With 35 billion total parameters and only 3 billion active per inference, it aims to break the serial bottleneck of autoregressive models through parallel decoding, marking a significant entry of diffusion models into high-performance AI agents.

Technically, Clover-35B demonstrates superior speed, generating at 680 tokens/s. In WindTunnel web agent evaluations, its median execution time is just 1.52 seconds—more than twice as fast as the Jev+Mercury combination—with higher success rates in code generation, tool calling, and complex task planning. These metrics confirm that the 'diffusion + MoE' approach can match or exceed autoregressive peers of similar size while offering substantial efficiency gains.

This development challenges the long-standing monopoly of autoregressive models on inference speed, providing developers with a more cost-effective foundation for real-time agents. By reducing latency in interactive AI applications, Clover-35B paves the way for broader industrial adoption of autonomous agents.

LATEST/Clover-35B is now fully open-source, with community focus shifting to its deployment in low-latency agent scenarios. Recent data confirms stable throughput of 680 tokens/s in complex code generation, while MoE sparse activation reduces memory footprint, offering a new engineering paradigm for deploying diffusion models on edge devices or high-concurrency servers.

Hourly heat curve

21 fully observed hours
Now
16.8
Peak
21.011:00
24h change
+42%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 2 reports · Latest first
  1. 量子位官网T2·85 pts

    Diffusion Intelligence Releases Clover-35B, Doubling Agent Speed Over Autoregressive Models

    • Clover-35B utilizes an MoE architecture with 35B total parameters and ~3B active per token.
    • Achieved a median execution time of 1.52s in WindTunnel web agent benchmarks, over 2x faster than competitors.
    • Successfully applies diffusion models to code generation, tool calling, and complex task planning, outperforming same-size autoregressive models.
  2. InfoQ 中文T2·85 pts

    Clover-35B Open-Sourced: Diffusion LLM Uses MoE to Challenge Agentic Coding

    • First introduction of MoE architecture into large-scale diffusion language models, balancing parameter scale and inference efficiency
    • Generation speed reaches 680 tokens/s, significantly outperforming traditional autoregressive serial decoding
    • Performance meets or exceeds same-size autoregressive models in certain agentic tasks and general benchmarks