Clover-35B-A3B-Preview
DiffuSpace's Apache-2.0 diffusion language model: 35B total, about 3B active parameters, block-parallel decoding on a Qwen3.6 backbone.
This entry is marked for review. Check the original sources before relying on its details.
Overview
Clover-35B-A3B-Preview is an open-weight diffusion language model from DiffuSpace, released around October 29, 2026 under Apache 2.0. It starts from Qwen3.6-35B-A3B (35B total, about 3B active parameters, 40 layers mixing Gated DeltaNet and full attention, 256 routed experts with top-8 routing) and adds bidirectional block attention so it can decode 32-token blocks in parallel, unmasking tokens once confidence passes 0.95. Context is 128K tokens and BF16 weights need about 69 GB of GPU memory. DiffuSpace claims state-of-the-art results among open diffusion language models; secondary reports cite 73.6 on SWE-bench Verified and 680 tokens/s, both company-reported, and one hands-on review found gaps versus Qwen on coding, multilingual output and logic puzzles.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 35B total / about 3B active MoE built on Qwen3.6-35B-A3B
- Block diffusion decoding: 32-token blocks generated in parallel, 0.95 confidence threshold
- 128K-token context window, text-only
- Apache 2.0 license; runs in Transformers, DiffuSpace's SGLang implementation and llama.cpp
- About 69 GB of GPU memory for BF16 weights
Use cases
- Research and benchmarking of diffusion language models against autoregressive ones
- Low-latency agentic coding experiments on 80 GB-class GPUs
- Serving with SGLang for throughput-oriented text generation
Pricing
Free open weights under Apache 2.0; you supply the 80 GB-class GPU or multi-GPU hardware.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 29 Oct 2026
- Record updated
- 2 Nov 2026
- Record last checked
- 2 Nov 2026