Skip to content
Neural Daily

ModelsNews

Clover-35B: open diffusion LLM hits 680 tok/s on one GPU

🇭🇺 Olvasd el magyarul →

DiffuSpace released Clover-35B-A3B-Preview, an open diffusion language model under Apache 2.0. It generates text in parallel blocks and, according to the company, beats same-size Qwen models on agentic coding.

By: Mikola Máté Patrik · · 2 min read

Bar charts of Clover-35B-A3B-Preview agentic benchmark results against LLaDA2.2 flash, Mercury 2.5 and Qwen models
Image: DiffuSpace · Press kit · Official benchmark chart from the DiffuSpace launch blog · Source

What was released?

DiffuSpace, a new company from the team behind HKU NLP's Dream 7B, released its first model, Clover-35B-A3B-Preview (DiffuSpace). Weights are available under Apache 2.0 on Hugging Face, GitHub and ModelScope.

A classic LLM writes token by token. Clover instead starts each 32-token block masked, fills the confident positions in parallel, and only then moves to the next block. That way, several tokens can be produced in one step.

Architecture

  • Derived from Qwen3.6-35B-A3B: 35B total, about 3B active parameters per token, 256 routed experts, top-8 routing.
  • Hybrid Gated DeltaNet and full attention, 40 layers, bidirectional attention inside each block.
  • 128K context, thinking mode, text only.

DiffuSpace's numbers

  • SWE-bench Verified: 73.6, Terminal-Bench 2.1: 51.7, τ³ Overall: 65.3. It beats Qwen3.6-35B-A3B, LLaDA2.2 flash and Mercury 2.5 on all six agentic tests.
  • GPQA Diamond: 83.8, ZebraLogic: 95.8, AA-LCR: 72.0.
  • Speed: 680 tok/s on one NVIDIA GB300 (BF16, batch 1, internal agent workloads). On an H200 it reaches 515 tok/s, against 365 tok/s for Qwen with DFlash2.

These are the company's own results and have not been independently verified.

What to keep in mind

BF16 weights need about 69 GB of GPU memory, and the transformers implementation requires trust_remote_code. For serving, the company recommends its own SGLang implementation, since the HF version can produce slightly different output. llama.cpp support comes with self-speculative decoding, but the backend code is not in the repo yet.

Why it matters

In agentic work, every generation delay adds up across the loop. If diffusion models really keep autoregressive quality at larger sizes, that means cheaper and faster self-hosted coding agents. Clover is the first open diffusion LLM that looks seriously competitive at this size.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles