DiffuSpace unveils Clover, an open-source model built for fast AI agents

DiffuSpace unveils Clover-35B-A3B, an open-source diffusion language model capable of reaching 680 tokens/s on a single NVIDIA GB300 GPU.

680 tokens generated per second on a single GPU, 35 billion parameters in total, and only three billion activated at each step. DiffuSpace unveils Clover-35B-A3B-Preview, a diffusion language model designed to accelerate text generation without sacrificing performance in reasoning, coding, and tool use.

Unlike traditional autoregressive models, which produce their response token by token, Clover leverages block-based generation that allows it to solve multiple text positions in parallel. The company claims that this approach enables it to outperform several models of comparable size on AI agent benchmarks.

The model weights are available under the Apache 2.0 license, with implementations for SGLang and llama.cpp. This initial version, still presented as a preview, can be downloaded and run locally. Another way to generate text

Most large language models use an autoregressive architecture. They successively predict each token based on those that have already been produced.

This operation creates a sequential dependency: the model must know the output of the previous step before continuing generation.

Diffusion language models, or dLLMs, explore a different approach.

Instead of generating all tokens in a strictly sequential order, Clover works on blocks of text whose positions are initially masked. The model estimates their content, then progressively refines the predictions during a denoising process.

Multiple positions can thus be solved during a single step, before the finalized block is added to the context available for the next generation.

This architecture makes it possible to further exploit GPU parallelism. It does not automatically guarantee speedups in all configurations, as denoising iterations and block sizes also influence performance.

DiffuSpace particularly highlights agentic applications, where generation delays accumulate across tool calls and correction loops. A 35-billion-parameter model derived from Qwen

Clover is based on a Mixture-of-Experts (MoE) architecture, in which only a portion of the network is mobilized to process each token.

The model gathers approximately 35 billion parameters, but activates only about three billion per token.

An important point appears in its Hugging Face model card: Clover is derived from Qwen3.6-35B-A3B, and not trained entirely from scratch.

DiffuSpace adopts its hybrid architecture combining Gated DeltaNet and attention mechanisms, before modifying it to support block-based diffusion generation.

The structure notably retains 40 layers and 256 routed experts, eight of which can be selected per token, in addition to one shared expert.

The changes mainly concern the bidirectional processing of information within blocks and the organization of previous token memory.

This lineage with Qwen is explicitly acknowledged by DiffuSpace, which thanks the Alibaba team for its work and training environment.

Clover is therefore not a simple rebranding of Qwen, but a specialized architectural adaptation of an existing base. This distinction is essential to understanding the nature of DiffuSpace's technical contribution. A new version of Gated DeltaNet for diffusion

At the core of the model, DiffuSpace modified the Gated DeltaNet (GDN) mechanism to better manage two types of information.

The first corresponds to the context already produced, preserved as a compact recurrent state. The second concerns the tokens in the block currently being generated, which must be able to interact with each other to improve their predictions.

The system combines compressed historical memory with bidirectional attention within the block.

The team also introduced a mechanism to calibrate the influence of this memory as the number of tokens to process increases. An additional convolution branch accounts for information located to the right of a position in the block.

These adaptations are intended to allow Clover to maintain the benefits of efficient sequential processing for previous content, while leveraging parallel generation within the current block.

The model also features a context window of 131,072 tokens, counting both the input and the generated response, and offers a reasoning mode. Up to 680 tokens per second on an NVIDIA GB300

The main performance figure put forward by DiffuSpace concerns generation speed.

In its internal measurements on agentic workloads, Clover reaches 680 tokens per second on a single NVIDIA GB300 GPU, in BF16 precision and with a batch size of 1.

On an NVIDIA H200, the lab announces 515 tokens per second, compared to 365 tokens per second for a Qwen configuration accelerated with DFlash2.

These results suggest a significant advantage for tasks requiring repeated responses, such as coding agents or systems using multiple tools.

However, a distinction must be made between generation throughput and the latency actually experienced by the user. Prompt processing time, tool execution, data transfers, and potential verification operations can alter the total time of a task.

The figures of 680 and 515 tokens per second come from DiffuSpace's tests. They do not yet constitute an independent, universal measure of the model's speed across all configurations. High results in coding and tool use

DiffuSpace presents Clover as particularly performant on agentic tasks.

In published evaluations, the model scores 73.6 on SWE-bench Verified, a benchmark dedicated to solving real software problems, and 51.7 on Terminal-Bench 2.1, which measures the ability to accomplish tasks using a terminal.

It also reaches 65.3 on τ³ Overall, dedicated to interactions and tool use.

According to DiffuSpace, Clover achieves the top results among selected models across the six agentic benchmarks in its comparison, notably ahead of several diffusion models and autoregressive models of similar size.

Announced general performance includes 83.8 on GPQA Diamond, 95.8 on ZebraLogic, and 84.8 on MMLU-Pro.

These figures show ambitious positioning for a diffusion model of this size. Their interpretation depends,