Diffusion language model scales to 35B parameters: Clover-35B-A3B-Preview beats a same-size autoregressive model on agent benchmarks and generates 680 tokens/s on one GPU
Synopsis
DiffuSpace released Clover-35B-A3B-Preview, a mixture-of-experts diffusion language model with 35 billion total parameters and 3 billion activated parameters, with a thinking mode. It asks whether diffusion language models stay competitive at larger scale while keeping the speed advantage of global generation; on agentic coding and tool-use benchmarks it outperforms the similarly sized autoregressive Qwen3.6-35B-A3B, and reaches 680 tokens/s on internal agent workloads on a single NVIDIA GB300 GPU with BF16 precision and batch size 1.
Interpretation
A diffusion language model scaled to 35 billion total parameters can still be compared head-to-head with a similarly sized autoregressive model while keeping faster generation, which is the answer Clover-35B-A3B-Preview gives. Dream 7B had shown that diffusion language models can go toe-to-toe with autoregressive models, but whether that holds at larger scale while preserving the speed advantage of global generation was a more fundamental question. Across multiple benchmarks, Clover-35B-A3B-Preview outperforms the similarly sized autoregressive Qwen3.6-35B-A3B while generating significantly faster, and it scores highest among the evaluated models on all six agent benchmarks in the comparison.
The model computes the readout of historical memory and the denoising inside the current block with different mechanisms, which suits block-level parallel denoising to long context. A gated DeltaNet processes a sequence causally, whereas block diffusion needs bidirectional interaction within the same block. Previously generated context is compressed into the gated DeltaNet recurrent state, tokens within the current block interact through fine-grained bidirectional softmax attention, the historical-memory readout joins that same softmax as a separate information unit, and learnable block-size calibration plus a zero-initialized right-side convolution branch are added.
On agent tasks, speed is itself a capability, because agents run through cycles of decision, execution, feedback and adjustment, and each generation delay accumulates across the pipeline. Faster generation shortens that feedback loop, so developers see application previews sooner and agents respond to execution results and move to the next step more quickly. On agent tasks on a single NVIDIA H200 GPU, Clover reaches 515 tokens/s against 365 tokens/s for the DFlash2 setup; in a full evaluation of 49 tasks with three attempts each, Clover's median agent execution time is 1.52 seconds against 8.86 seconds for Gemini3.6 Flash.
Perspective
The demonstrated scope is a mixture-of-experts diffusion language model with 35 billion total and 3 billion activated parameters, evaluated on public benchmarks and internal agent workloads, plus an INT4-quantized on-device deployment on an NVIDIA RTX 5070 Ti with 16GB of VRAM. The intended readers are engineers and researchers tracking language-model architecture and agent inference cost. The releasing team says it will continue scaling diffusion language models and improving multimodal capability, long-horizon task stability and inference efficiency.
The speed figures come from internal agent workloads on specified hardware, including a single NVIDIA GB300 GPU with BF16 and batch size 1 and a comparison against a DFlash2-accelerated setup on a single NVIDIA H200 GPU, not from a cross-platform general measurement. The decision model built on this model is not yet released and is planned for a later date, and the agent execution harness and desktop coding client are still under development.
