Key Findings
Clover-35B-A3B-Preview demonstrates significant performance gains in agent-centric tasks, recording a score of 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1, and 65.3 on τ³ Overall. These results surpass those of autoregressive models Qwen3.6-35B-A3B and Qwen3.5-35B-A3B, as well as diffusion models LLaDA2.2 flash and Mercury 2.5. In local testing using a GeForce RTX 5070 Ti (INT4 quantization), the model reached 101.8 tok/s on MATH-500, roughly double the 48.5 tok/s recorded by similar Qwen and MiniCPM models.
Technical Details
Unlike standard LLMs that generate text token-by-token, Clover utilizes a diffusion language model architecture to generate text in blocks. The model is a Mixture-of-Experts (MoE) type with 35 billion total parameters and 3 billion active parameters. It supports a maximum context window of 131,072 tokens (128K) and includes a thinking mode. The model is released under the Apache 2.0 license as open weights.
- Hardware Performance: On a single NVIDIA GB300 (BF16, batch size 1), it reached 680 tok/s. On a single NVIDIA H200, it achieved 515 tok/s, exceeding the 365 tok/s of a Qwen model accelerated by DFlash2.
- General Benchmarks: The model scored 83.8 on GPQA Diamond, 95.8 on ZebraLogic, and 72.0 on AA-LCR.
- Compatibility: Inference code is available via Hugging Face, with future support for SGLang and llama.cpp to be published on GitHub.
Background and Context
DiffuSpace is a development organization formed by members of the team that created Dream 7B, a diffusion language model released in April 2025 by the University of Hong Kong’s NLP group. The Dream series has accumulated over 2.7 million downloads on Hugging Face.
Significance and Outlook
The model is specifically designed for text-only input focused on agent-type coding and tool utilization. DiffuSpace is currently developing a desktop coding client and an agent execution harness to integrate natural language instructions, tool calling, task execution, and result inspection into a unified workflow.
Source: https://news.yahoo.co.jp/articles/7da4f72ce31bd95fef2a44e6176e6a59e99a646c
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
