Clover-35B-A3B-Preview

Developers can run it locally to generate text and conversational responses using the provided inference implementations.

Availability

Available on Hugging Face

Pricing model

Not established

Deployment

Standalone PyTorch inference; GPU with enough memory for BF16 weights (approximately 69 GB) plus inference state

Privacy

Not established

Licence

Apache-2.0

• The model is available on Hugging Face. ^[1]

  The model implementation derives from the Hugging Face Transformers Qwen3.5-MoE implementation used for the Qwen3.6 architecture, Copyright 2025 The Qwen Team and The HuggingFace Inc.

• Clover combines a 35B-parameter mixture-of-experts backbone with approximately 3B active parameters per token. ^[1]

  Clover combines a 35B-parameter mixture-of-experts backbone with approximately 3B active parameters per token, block diffusion generation, and a **128K context window**.

• The model supports up to 131,072 tokens for the prompt and generated response together. ^[1]

  - **128K context:** up to 131,072 tokens for the prompt and generated response together.

• The Hugging Face implementation uses block diffusion with LowConfidence decoding. ^[1]

  - **Block diffusion with LowConfidence decoding:** this Hugging Face implementation starts each block with masked tokens and resolves confident positions in parallel.

• The quick-start instructions require a GPU with enough memory for the BF16 weights (approximately 69 GB) plus inference state. ^[1]

  Use a GPU with enough memory for the BF16 weights (approximately 69 GB) plus inference state.

• DiffuSpace recommends its official SGLang implementation for serving and evaluation. ^[1]

  **We recommend our official SGLang implementation for serving and evaluation.** Setup instructions and supported decoding modes are available in the [official repository](https://github.com/DiffuSpace/Clover-35B-A3B-Preview).