DiffuSpace Releases Clover-35B-A3B-Preview Coding Model
The 35-billion-parameter open-source model uses a mixture-of-experts architecture to accelerate code generation.
Updated on Oct. 29, 2026 in Artificial Intelligence
Live Poll
Do you trust open-source artificial intelligence tools to handle your professional coding or research tasks?
DiffuSpace has released its Clover-35B-A3B-Preview model as an open-source tool under the Apache 2.0 license. This 35-billion-parameter model uses a mixture-of-experts architecture and is currently available for download on Hugging Face, GitHub, and ModelScope.
Why it matters
The model demonstrates a shift toward high-efficiency inference in coding tasks by utilizing a mixture-of-experts design that reduces computational overhead. This release provides developers with a new open-source alternative for automated software engineering workflows.
Clover-35B-A3B-Preview achieves 515 tokens per second on an NVIDIA H200 while using only 3 billion active parameters per token. It scored 65.3 on the tau-cubed overall benchmark and 67.9 on BFCL-v4.
The players
DiffuSpace
An artificial intelligence developer focused on open-source model releases and software engineering agents.
NVIDIA
A semiconductor manufacturer and leading supplier of H200 high-performance computing hardware used for AI model inference.
The details
The model employs a Gated DeltaNet backbone, a neural network architecture designed for sequential data, coupled with learnable block-size calibration to optimize processing. It uses block diffusion—a generative technique that constructs output in distinct, sequential blocks—and performs parallel denoising within those blocks to synthesize tokens more efficiently than standard transformer models.
Timeline
October 29, 2026: DiffuSpace released the Clover-35B-A3B-Preview model.
The Tech Race
This release follows the trend of open-source models aiming to climb the leaderboard of the SWE-bench Verified benchmark. DiffuSpace is positioning this model as a lightweight alternative to larger, denser coding models through its mixture-of-experts efficiency.
Developers can download the model weights from Hugging Face, GitHub, or ModelScope to integrate the code-generation capability into existing pipelines. A dedicated desktop coding client and agent harness remain in development for future release.
The takeaway
The Clover-35B-A3B-Preview represents a technical pivot toward efficient, block-based generation for coding tasks. Watch for the forthcoming release of DiffuSpace’s desktop coding client, which will serve as the first practical test for this model's agentic capabilities.
Further reading
For more on the current state of software-coding AI, see our latest coverage in Artificial Intelligence.
Source note: This article includes information reported by Pandaily.
Live Poll
Do you trust open-source artificial intelligence tools to handle your professional coding or research tasks?
