A diffusion model at 680 tokens/s on a single GPU—don’t rush to treat it as a replacement

Clover-35B-A3B-Preview has been released. DiffuSpace labels it plainly: the first open-weight production-grade diffusion language model, 35B total parameters, 3B active, Apache 2.0, running at 680 tokens/s in BF16 on a single GPU.

Set the numbers aside for now. The question diffusion language models get asked most often is: apart from paper figures, where can it actually run? This time Clover releases the weights, inference backends, and demos together, so at least that question now has an executable answer.

Autoregressive models write token by token, while diffusion models can denoise a block of tokens together. Clover uses block diffusion: officially, it compresses historical context into a GDN recurrent state, uses bidirectional softmax attention within the current block for fine-grained interaction, and adds learnable block-size calibration plus a zero-initialized convolutional branch for right-side context. In plain terms: it wants block-level parallel denoising to work in long contexts while preserving generation speed.

Across the official benchmark results, Clover is highest among the listed models on six items: SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7, and τ³ Overall 65.3. On general capabilities, GPQA Diamond 83.8, ZebraLogic 95.8, AA-LCR 72.0. Comparisons include same-size Qwen3.6-35B-A3B as well as diffusion models such as LLaDA2.2 flash and Mercury 2.5.

Pay attention to the conditions behind the speed data. The 680 tokens/s figure was measured on a single NVIDIA GB300, BF16, batch size = 1, using internal Agent-scenario data. The official release also gives Agent-task data on H200: Clover 515 tokens/s, while the DFlash2 solution reaches 365 tokens/s. This is not a general SLA, but it shows the speed potential of the diffusion architecture in single-request Agent loops.

Speed in agents is not just benchmark decoration. An Agent must continuously make decisions, execute, get feedback, and adjust; each step's latency accumulates along the task chain. In the official demo, for the same GitLab task, with the same Agent framework, prompts, tools, and browser, Clover completed creating a private project and inviting members in 13 steps and 169.9 seconds, while Qwen3.6 Flash took 21 steps and 392.3 seconds.

Another demo is route planning. The user wants to drive from CMU in Pittsburgh to the nearest Wendy's. Map search has no location signal, and directly searching the store name jumps to Maine. Clover recognizes the result has deviated from the target and progressively adds city, neighborhood, and ZIP code to rewrite the search; after the page gets stuck, it resets and re-enters, finally finding a store 2.0 km away, about a 3-minute drive. In the official record, Clover finished in 28 steps and 169.9 seconds, while Qwen3.6 Flash used all 30 steps and 578.3 seconds and still stopped at a store 347 km away.

These demos have small samples, and the official release also states they are a single run with no retries. They can show behavior, but they cannot replace independent evaluation. What really matters is whether the speed and success rate still hold after you switch to your own toolchain, your web environment, and your long-running tasks.

Officially, there is also a drone scheduling case. A composite air-traffic-control message needs to be split into five tasks: no-fly zone cancellation, waypoint modification, task assignment, no-fly zone radius adjustment, and emergency scheduling. Clover gives all five correct results simultaneously after 0.97 seconds, while Qwen shows results in 3.78 seconds, misparsing UAV5's mapping task as a medical task. Here you can see the appeal of parallel denoising for parsing composite instructions.

GenUI is another natural scenario. The official demo uses a single HTML file to simulate a desktop environment, requiring draggable windows and small apps such as Notepad, Calculator, Clock, Weather, and Settings. Clover generates an interactive page with a dock and window system from a single requirement.

There is also on-device data. Clover is deployed with INT4 quantization on an NVIDIA RTX 5070 Ti with 16GB of VRAM, achieving 101.8 tokens/s generation speed on MATH-500; the official comparison with same-scale models such as MiniCPM and Qwen is 48.5 tokens/s. This number is more useful as a reference for people who want to run Agents or interactive applications locally.

Ecosystem partner SigmaZ brings Clover into education scenarios for interactive 3D science popularization. Officially, it shows a cell energy factory page where users can click organelles, switch tabs, and rotate the view. Another is a solar and lunar eclipse demo where you drag the timeline to see the relative motion of the Sun, Earth, and Moon. SigmaZ says it plans to increase token throughput to 2,000 tokens per second.

Officially, a decision model based on Clover is also previewed for a later release. It supports independent mode and joint mode; joint mode puts multiple questions under the same material into one denoising process and generates them in parallel, with up to 8 answers each time. Internal validation includes 500 groups and 2,420 related questions. In joint mode, per-question accuracy is 95.0%, and 80.4% of question groups are all correct; the same model answering independently gets 89.8% and 62.4%, while TypeSafe Jev in historical evaluations gets 63.6% and 15.8%. This direction is more product-relevant than simply chasing generation speed: generating multiple related judgments together may reduce contradictions.

How to use it. Weights are on Hugging Face and ModelScope, code is on GitHub, Apache 2.0. Officially supported: Hugging Face Transformers, SGLang, and llama.cpp. SGLang supports block diffusion denoising and self-speculative decoding; llama.cpp supports self-speculative decoding. The official recommendation is to use SGLang for serving and evaluation, because there are numerical differences between the Hugging Face implementation and the SGLang kernel, and the generated text may differ. The minimal Transformers example requires PyTorch 2.10 or above and Transformers 5.19.0; BF16 weights on a single GPU are about 69GB, so leave enough VRAM. SFT code has not been released yet; the official note says coming soon.

Judging from the public materials, Clover pushes diffusion language models one step forward: it no longer only proves that it can compete, but puts open weights, inference backends, Agent demos, and on-device deployment on the same table. What to watch next is not another official benchmark score, but independent reproduction, long-horizon task stability, and server-side concurrency and cost. 680 tokens/s is impressive, but before putting it into a production pipeline, first get SGLang running, then test it against your real tasks.

Related links: Official blog Hugging Face ModelScope GitHub Online demo application SGLang PR llama.cpp PR