Clover-35B-A3B-Preview: A New Way for Intelligence
Extraordinary things are never achieved by following conventional paths.
Today, we are officially releasing Clover-35B-A3B-Preview, the first version of DiffuSpace Clover (三叶).
In April 2025, the NLP Group at The University of Hong Kong released Dream 7B. To date, the Dream family of models has accumulated more than 2.7 million downloads on Hugging Face. Building on Dream, members of the team began a new chapter at DiffuSpace, while remaining committed to making AI broadly accessible through open source.
If Dream 7B proved that diffusion language models can go toe-to-toe with autoregressive models, Clover sets out to answer a more fundamental question: When scaled up, can diffusion language models maintain their competitiveness while preserving the speed advantage of global generation?
Clover-35B-A3B-Preview provides a clear answer: yes. At a larger scale, diffusion language models can even edge out similarly sized autoregressive models while delivering significantly faster generation. Speed and performance can coexist in a single model.
Clover-35B-A3B-Preview is a Mixture-of-Experts (MoE) diffusion language model with 35 billion total parameters and 3 billion activated parameters, with support for a thinking mode. Across multiple benchmarks, it outperforms the similarly sized autoregressive model Qwen3.6-35B-A3B, while generating significantly faster.
Clover-35B-A3B-Preview particularly excels in agentic coding and tool-use tasks. Across all six agent benchmarks in our comparison, it achieves the highest scores among the evaluated models. Notably, it scores 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1, and 65.3 on τ³ Overall, outperforming fellow diffusion language models LLaDA2.2 flash and Mercury 2.5, as well as similarly sized autoregressive models Qwen3.6-35B-A3B and Qwen3.5-35B-A3B.
Clover-35B-A3B-Preview also delivers strong general-purpose capabilities. Across benchmarks covering instruction following, knowledge, logical reasoning, and long-context understanding, it matches or exceeds the performance of similarly sized autoregressive models, achieving 83.8 on GPQA Diamond, 95.8 on ZebraLogic, and 72.0 on AA-LCR.
In terms of inference efficiency, Clover-35B-A3B-Preview achieves an overall generation speed of 680 tokens/s on internal agent workloads, tested on a single NVIDIA GB300 GPU with BF16 precision and a batch size of 1. Even compared with a similarly sized Qwen model accelerated by DFlash2, Clover offers a better balance between performance and inference speed. On a single NVIDIA H200 GPU, Clover also reaches 515 tokens/s on agent tasks, compared with 365 tokens/s for the DFlash2 setup. These results demonstrate the speed advantage of the diffusion architecture in complex agent workloads.
Building on Clover-35B-A3B-Preview, we will continue to advance the scaling of diffusion language models, improve model capabilities and inference speed, and validate and refine their performance across a broader range of real-world tasks.
Clover-35B-A3B-Preview is now open source under an Apache license. Model weights and related code are available on Hugging Face and GitHub, with inference support for SGLang and llama.cpp. Developers can download and deploy the model, build applications, and conduct further research on top of Clover. We look forward to working with the community to push the boundaries of what diffusion language models can achieve.
Related Links:
HuggingFace：https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview
GitHub：https://github.com/DiffuSpace/Clover-35B-A3B-Preview
ModelScope：https://modelscope.cn/models/Diffuspace/Clover-35B-A3B-Preview
Request Online Access：https://chat\.diffuspace\.ai/
Diffusion: Pushing Toward Greater Possibilities
The leap forward is driven by more than parameter scaling. It also comes from our architectural redesign of the Gated DeltaNet (GDN) for Block Diffusion.
We redesigned the Bidirectional Gated DeltaNet to handle historical memory and intra-block denoising using different computational mechanisms. Previously generated context is efficiently compressed into GDN's recurrent state, while tokens within the current block interact through fine-grained bidirectional softmax attention.
Crucially, we treat the readout from historical memory as a separate information unit that participates in the same softmax operation as the tokens within the current block. This allows the model to learn when to rely on historical context and when to draw on information within the current block. We also introduce learnable block-size calibration to compensate for the natural dilution of attention to historical memory as the number of tokens within a block increases. For local feature extraction, we add a separate, zero-initialized convolution branch for the right-side context. Starting from the pretrained causal convolution behavior, the model can directly learn distinct features from both left and right contexts.
Together, these designs integrate compact long-range memory, selective intra-block interactions, and bidirectional local context modeling, making the GDN architecture better suited to block-level parallel denoising.
The speed and performance of Clover-35B-A3B-Preview stem directly from these targeted architectural innovations.
Agentic Coding
Clover-35B-A3B-Preview goes beyond code completion. It can translate natural-language requirements directly into complete, interactive applications, while also understanding webpages and executing multi-step operations in real browser environments.
For AI agents, speed is itself a core capability. Agents operate through continuous cycles of decision-making, execution, feedback, and adjustment. Every generation delay accumulates across the task execution pipeline.
Clover's faster generation speeds shorten this feedback loop. Developers can see application previews sooner, while agents can respond to execution results more quickly, adapt their strategies, and move on to the next step.
When every step becomes faster, agents can explore more possible paths within the same amount of time and reach results sooner.
On-Device Applications
Edge computing demands rapid, holistic judgment rather than "slow thinking."
For on-device deployment, Clover has been successfully deployed with INT4 quantization on an NVIDIA RTX 5070 Ti with just 16GB of VRAM. On MATH-500, Clover achieves a generation speed of 101.8 tokens/s, demonstrating a substantial efficiency advantage over the 48.5 tokens/s recorded in comparison tests with similarly sized MiniCPM and Qwen models. We will continue optimizing the inference engine and on-device deployment setup to further unlock the performance of diffusion language models on resource-constrained devices.
We compared Clover-35B-A3B-Preview against representative lightweight, inference-efficient, and on-device-deployable models, including MiniCPM-V 4.7 (35B-A3B), LFM2.5 (24B-A2B), Nemotron 3.5 Lightning (30B-A3B), and Gemma 4 (26B-A4B).
Clover achieved the highest scores on six out of nine benchmarks, demonstrating advantages across knowledge reasoning, tool use, and agentic coding.
Notably, Clover scores 51.7 on Terminal-Bench 2.1 and 73.6 on SWE-bench Verified, outperforming MiniCPM-V 4.7, which shares the same 35B-A3B parameter configuration, by 15.7 and 28.0 points, respectively.
For tool use, Clover achieves 67.9 on BFCL-v4 and 14.4 on τ³ Banking. For knowledge reasoning, it scores 83.8 on GPQA Diamond and 84.8 on MMLU-Pro, leading all other models included in the comparison on these benchmarks.
Demo showcases
Route Planning
In this example, the user asks"I am at CMU Pittsburgh，how long it takes to drive to the neareat Wendy's?"
However, mapping websites lack spatial location signals, meaning a direct search for the store name incorrectly defaults to Portland, Maine. Clover identifies that the result deviates from the target, progressively refining the search by adding city, neighborhood, and zip code keywords. When the page stalls, it resets and refills the route twice, eventually locating the target store and successfully retrieving a driving distance of 2.0 km and a travel time of roughly 3 minutes.
In this comparison, Clover-35B-A3B-Preview completed the task in 28 steps over 169.9hausted all 30 steps and took 578.3 seconds, yet remained stuck at a store 347 km away (approximately 4.5 hours of driving time from the starting point), failing the evaluation criteria.
WebMCP: Completing Six Business Tasks in Sequence
Using tools exposed through webpages, the model completes a sequence of business tasks, including invoice creation, course registration, amount aggregation, event status checks, and customer information retrieval.
The demonstration places multiple business webpages within a single view, comparing tool calls, webpage interactions, and task outcomes. Both models successfully complete the six selected tasks.
In the full evaluation covering 49 tasks, with three attempts per task, Clover achieves a median agent execution time of 1.52 seconds, compared with 8.86 seconds for Gemini3.6 Flash.
Creating a Private Project and Inviting Members
Based on a single prompt, the model is required to create a private project named "planner" and add two designated members. This requires sequential execution of project creation, visibility settings, member searching, and invitations, while consistently maintaining alignment with the user's objective.
Under identical agent frameworks, prompts, tools, and browser configurations, Clover-35B-A3B-Preview completed the task in 13 steps and 169.9 seconds, whereas Qwen3.6 Flash took 21 steps and 392.3 seconds.
GenUI: Building an Interactive Desktop in a Single HTML File
In this example, the user asks the model to create a simulated desktop environment using a single HTML file, complete with draggable windows and multiple functional applications.
Clover generates an interactive desktop featuring a dock, a window management system, and applications including a notepad, calculator, clock, weather app, and settings panel.
In the demo, users can open apps, input text, operate the calculator, and adjust settings. This task rigorously tests interface layout, interactive logic, and code organization across multiple functional modules, showcasing the generation pipeline from natural language requirements to fully interactive applications.
UAV Dispatch: Parsing Compound ATC Messages in Parallel
n this urban low-altitude UAV dispatch demo, the model parses a compound air traffic control (ATC) message into multiple tasks, including lifting no-fly restrictions, modifying waypoints, assigning missions, adjusting no-fly zone radii, and emergency dispatch. Clover produces all five correct results together in 0.97 seconds. Qwen presents its results after 3.78 seconds, but incorrectly interprets UAV5's mapping mission as a medical mission. This case demonstrates the advantages of diffusion-based parallel generation in both response speed and task intent understanding when parsing compound instructions.
Ecosystem Partner
SigmaZ is our exclusive ecosystem partner for GenUI. They are bringing Clover into education, transforming static text and images into visual experiences that users can explore and interact with. SigmaZ has launched a closed beta demo for educational use. Apply for access at edu.tap8.ai. We are now beginning joint post-training, aiming to increase token throughput to 2,000 tokens per second and extend real-time visual coding to a broader range of interactive applications.
3D Science Exploration: Inside the Cell's Energy Factory
To explain how cells produce energy, the model builds an interactive 3D explainer with a cutaway view of an animal cell. Users can click organelles to move closer and read explanations, or toggle structural labels, the cutaway view, and rotation to explore the cell from the whole down to its individual components. By bringing knowledge explanations, a 3D scene, and interactive logic together on one page, the demo turns the conversion of glucose into cellular energy from a written description into a process users can understand through observation and interaction.
Interactive Astronomy: Visualizing How Eclipses Happen
To explain why solar and lunar eclipses do not occur every month, the model builds a time-driven 3D demonstration of the Sun, Earth, and Moon. It brings celestial positions, the tilt of the Moon's orbit, and changing lunar phases into a single scene, making the special alignments required for eclipses visible.
Users can drag a timeline to observe the bodies' relative motion, toggle orbits, labels, and shadow cones, or jump to a solar eclipse to see how the Moon's shadow falls on Earth. The page also notes where distances and sizes have been adjusted, balancing visual clarity with an accurate presentation of the underlying scientific concepts.
Decision Model
We are also introducing a decision model based on Clover-35B-A3B-Preview for fast, structured judgment tasks. It will be released at a later date.
The model supports two response modes:
Independent Mode: The model is called separately for each question, answering one question at a time. This mode is suited to standalone or routine judgments that require a fast response.
Joint Mode: Multiple questions about the same material are grouped together, and their answers are generated in parallel within a single denoising process, with up to eight answers per call.
What sets this model apart from TypeSafe Jev and other open-source decision models is Joint Mode's coordinated handling of related decisions. Rather than treating each decision in isolation, it substantially reduces the time spent on repeated interactions and allows judgments to inform one another during generation, helping avoid logical conflicts. Making multiple related decisions together is not only faster, but also more accurate.
In an internal validation set containing 500 groups and 2,420 related questions, Joint Mode answered 95.0% of individual questions correctly and got every answer right in 80.4% of question groups. In comparison, the same model in Independent Mode scored 89.8% and 62.4%, respectively. TypeSafe's decision model Jev scored 63.6% and 15.8% in historical evaluations on the same questions. In this validation set, processing related questions together improved both individual-question accuracy and the share of groups with all answers correct.
The Same Contract, Five Correlated Judgments
Rules within the same contract regarding "whether a notice was timely" and "whether a security deposit should be retained" share a logical dependency. When judged independently question by question, a model might simultaneously conclude that a notice was overdue while selecting the "retain deposit" option, which is only permissible if the notice is timely. Joint Mode enables these judgments to be generated concurrently, helping to minimize inconsistencies while reducing the overhead of repeatedly reading material and executing sequential calls.
In this validation case, Joint Mode answered all 5 judgments correctly, while Independent Mode answered 4 correctly, and Jev answered 2 correctly. The demo compares judgments on notice deadlines and deposit retention, demonstrating how Joint Mode handles mutually dependent rules.
Chinese Chess: Making Sequential Decisions from the Board State
The decision model reads the current board state, selects its next move, and outputs a structured move action. In this game against TypeSafe Jev, DiffuSpace's decision model played Black and won.
Next Steps
Moving forward, we will continue pushing the boundaries of diffusion language model scaling, expanding multimodal capabilities, enhancing stability in long-horizon tasks, optimizing inference efficiency, and advancing quantization and edge-device adaptation.
At the same time, we are developing an agent execution harness and a desktop coding client to bring natural-language instructions, tool calls, task execution, and result inspection into a more integrated workflow. These tools are currently under development and will be rolled out shortly.
None of this progress would be possible without the foundations laid by the open-source community. We thank Qwen for its open model architectures and training ecosystem, which provided a reliable starting point for the research behind Clover-35B-A3B-Preview. We also thank the SGLang community for its engineering support for block diffusion inference, and the llama.cpp community for its open integration interfaces and collaborative development process.
We invite you to visit diffuspace.ai to apply for access to Clover-35B-A3B-Preview.
Developers can also download model weights from Hugging Face and ModelScope, build applications, or contribute through GitHub Issues and integration efforts. Feedback from real-world tasks will help us continuously improve Clover. We look forward to building together with the open-source community, and we believe diffusion models will be better for it.
We also welcome ecosystem partners across a range of use cases to bring diffusion language models into real-world tasks, working together to validate and refine them.
"Clover" stands for good luck, while carrying our highest aspirations for diffusion models.
Today, Clover breaks new ground, ready to grow across the vast frontier of intelligence.
