DiffuSpace open-sourced Clover-35B-A3B-Preview, saying GB300 a single GPU can reach 680 Token/s

Article

DiffuSpace changed Qwen3.6 to a diffusion model,GB300 running on a single GPU 680 Token/s Dongcha Beating AI flash: Diffusion Intelligence (DiffuSpace), founded by University of Hong Kong professor Kong Lingpeng, has released and open-sourced Clover-35B-A3B-Preview.

The model is based on a modification of Alibaba's Qwen3.6-35B-A3B architecture, with a total of 35000000000 parameters and approximately 3000000000 parameters activated per inference.

Traditional large models usually generate tokens one by one from left to right, whereas Clover uses block diffusion: it generates a short segment of text at a time, then completes the output through parallel prediction and repeated refinement. The team also modified the Gated DeltaNet, compressing and storing historical information while enabling bidirectional interaction among tokens within the current text block, balancing generation speed and contextual understanding.

In the six Agent evaluations officially disclosed, Clover achieved the highest score among the listed models in every case.SWE-bench Verified reached 73.6%，Terminal-Bench 2.1 reached 51.7%, surpassing the original Qwen3.6-35B-A3B, as well as LLaDA2.2 Flash、Mercury 2.5 and other diffusion models. However, the comparison did not include Mercury October 19 released by 3.

In terms of speed, on a single NVIDIA GB300, Clover reached a generation speed of BF16 precision, running an internal Agent task with single concurrency. On 680 Token/s it reached H200, exceeding the same-size Qwen model accelerated with 515 Token/s(DFlash2). The team also deployed the 365 Token/s quantized version on a INT4 with only 16GB of VRAM, achieving a speed of RTX 5070 Ti, with MATH-500 testing reaching 101.8 Token/s.

Original link https://m.theblockbeats.info/flash/374154

Key points

DiffuSpace released and open-sourced Clover-35B-A3B-Preview。
Clover is based on a modification of Alibaba's Qwen3.6-35B-A3B architecture, with a total of 35000000000 parameters and approximately 3000000000 parameters activated per inference.
Clover uses block diffusion to generate text and also modified the Gated DeltaNet。
In the six Agent evaluations officially disclosed, Clover achieved the highest score among the listed models in every case.
The team's disclosed internal tests showed that, on a single NVIDIA GB300 with BF16 precision, when running an internal Agent task with single concurrency, Clover's generation speed reached 680 Token/s。
The team also disclosed that Clover reached H200 on 515 Token/s, and deployed the INT4 quantized version on a 16GB with RTX 5070 of VRAM.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Traditional large models usually generate tokens one by one; Clover first generates a short segment of text, then completes the output through parallel prediction and repeated refinement. It also compresses and stores historical information and enables bidirectional interaction among tokens within the current text block.

Why it matters to readers

This indicates that diffusion models may increase generation speed for certain Agent tasks and specific hardware configurations, but the available data primarily consists of test results disclosed by the team.

Token/s refers to the number of tokens generated per second; a higher value usually means faster output;BF16 and INT4 are different numerical precisions and quantization methods that affect speed, VRAM usage, and performance.

Risks and unknowns

The performance and evaluation results mainly come from disclosures by the team, and independent reproduction is still lacking.
The evidence does not provide complete test configurations, model weights, or code information.
The cross-model comparison list did not include October 19 Mercury released by 3。
Speeds measured under different hardware, precision, concurrency, and task settings cannot be directly compared as equivalent.

DiffuSpace open-sourced Clover-35B-A3B-Preview, saying GB300 a single GPU can reach 680 Token/s

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center