Clover-35B-A3B-Preview Review: DiffuSpace Open-Weight MoE Model Capabilities and Access Options

Clover-35B-A3B-Preview delivers a 35-billion-parameter open-weight language model from DiffuSpace that uses block diffusion instead of standard autoregressive generation. Announced in late October 2026, the model targets agentic coding and tool-use workloads while maintaining competitive benchmark scores on tasks such as SWE-bench Verified. Its Apache-2.0 license and Hugging Face release make the weights immediately available for local or cloud deployment.

Clover-35B-A3B-Preview review: architecture, generation method and reported performance

The model combines a hybrid Gated DeltaNet and full-attention MoE backbone across 40 layers with 256 routed experts and top-8 routing plus one shared expert. It supports bidirectional attention within fixed-size blocks and generates text through iterative denoising rather than left-to-right token prediction. Default settings use a block size of 32 and LowConfidence decoding with a 0.95 threshold, while an optional thinking mode allows explicit reasoning steps before final output.

DiffuSpace reports 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1, and 65.3 overall on τ³, exceeding the base Qwen3.6-35B-A3B on the same agent benchmarks. General capabilities include 83.8 on GPQA Diamond and 95.8 on ZebraLogic. Inference speeds reach 680 tokens per second on a single NVIDIA GB300 in BF16 for internal agent workloads and 101.8 tokens per second in INT4 on an RTX 5070 Ti during MATH-500 evaluation.

Context length extends to 128K tokens. The release includes quick-start code for the Transformers library that requires trust_remote_code=True and PyTorch 2.10 or newer with CUDA. SGLang and llama.cpp integration notes appear in the repository, and an online demo is available at chat.diffuspace.ai after account approval.

API access, pricing status, deployment requirements and comparison points

DiffuSpace provides an OpenAI-compatible endpoint at api.diffuspace.ai/v1 under the model ID Clover-35B-A3B-Preview. Users must create an account, complete verification, and request an API key; no public pricing table has been published as of November 2026. Weights are hosted on Hugging Face and ModelScope at approximately 69 GB in BF16 format.

Deployment requires a CUDA-capable GPU for practical speeds, though the 3 billion active parameters per token allow the model to run on consumer hardware in quantized form. No formal security audit or license restrictions beyond Apache-2.0 are documented, and the model remains text-only without multimodal support.

Limitations include the preview status with SFT and fine-tuning scripts listed as forthcoming, reliance on provider-reported benchmarks, and the need for account approval to access both the online demo and API. Community contributions through GitHub and Hugging Face remain the primary route for early validation and optimization.