Clover 35B-A3B opens diffusion language weights for coding agents

DiffuSpace has released Clover-35B-A3B-Preview, a downloadable language-model checkpoint that uses block diffusion rather than ordinary token-by-token generation for coding and agent tasks. The published mixture-of-experts architecture has 35 billion total parameters and roughly three billion active parameters per token. Its official Hugging Face repository lists Safetensors weights and Apache-2.0 terms. That makes the release an inspectable piece of the growing alternative-decoding ecosystem, rather than a speculative architectural slide.

The model builds on Qwen3.6-35B-A3B and changes how the hybrid Gated DeltaNet and attention layers handle generation blocks. Tokens inside a block are refined in parallel, while previous blocks remain fixed as context. The official model card describes a combined context window of 131,072 tokens and default block-generation controls. These details determine what implementation is being tested; a conventional autoregressive serving configuration cannot be assumed to reproduce its latency or output behavior.

DiffuSpace reports a SWE-bench Verified score of 73.6, Terminal-Bench 2.1 at 51.7 and an aggregate agent score of 65.3. The announcement also describes 680 generated tokens per second on a single NVIDIA GB300, using BF16 and batch size one for internal agent workloads. All of those numbers are the developer's own results. Open Model Weights has not reproduced the tests, and throughput measured in one decoding setup should not be confused with reliable task completion or universal cost savings.

The source loading instructions call for trust_remote_code and approximately 69 GB of BF16 weight memory before additional inference state. Developers should review custom code, pin the repository revision and test the actual hardware and backend, including SGLang or llama.cpp integrations. Clover offers an intriguing open-weight route to faster agent loops, but there is not yet independent evidence that it outperforms all comparable models across ordinary workloads, safety constraints and deployment costs.