Back home

Clover-35B

x+web+hf+github 0.92
Visit source

Intro

Clover-35B (full name Clover-35B-A3B-Preview, also branded Clover / 三叶) is the first production-oriented diffusion language model (dLLM) released and open-sourced by DiffuSpace on 2026-10-28. It is a Mixture-of-Experts model with about 35B total parameters and ~3B active per token, block-diffusion decoding, up to 128K context, and an Apache-2.0 license.

The team traces back to HKU NLP's Dream 7B line. Weights and docs shipped together via the company blog, GitHub, Hugging Face / ModelScope, and @DiffuSpace_Labs on X; Chinese tech press (LeiPhone, QbitAI) covered the launch the same day.

Blog: diffuspace.ai/blog/clover-35b · GitHub: DiffuSpace/Clover-35B-A3B-Preview · HF: Diffuspace/Clover-35B-A3B-Preview · ModelScope: Diffuspace/Clover-35B-A3B-Preview.

Explain

Key points

  1. Paradigm: instead of token-by-token autoregression, Clover fills/denoises tokens in parallel inside blocks, then advances across blocks. DiffuSpace positions it for agentic coding and tool use, with a thinking mode.
  2. Vendor-reported numbers: SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7, τ³ Overall 65.3, GPQA Diamond 83.8; ~680 tokens/s on one NVIDIA GB300 (BF16, bs=1) on internal agent workloads, and ~515 tokens/s on H200 (vs ~365 for a DFlash2-accelerated same-size Qwen setup in their write-up).
  3. Architecture: Bidirectional Gated DeltaNet tuned for block diffusion, learnable block-size calibration, and a zero-initialized right-context convolution branch; inference backends called out are SGLang and llama.cpp.
  4. On-device: INT4 on an RTX 5070 Ti with 16GB VRAM is claimed; MATH-500 self-report ~101.8 tok/s.
  5. Roadmap: a decision model (independent vs joint mode, up to 8 related questions in one denoising pass), multimodal follow-ons, plus an agent harness and desktop client — decision weights are not the Preview drop itself.

Why this radar term

Clearest named model launch in this window: after the official X post (~03:42 UTC / 11:42 Asia/Shanghai), afternoon Chinese coverage and developer chatter spiked, and earlier radar runs had not stored the term. Concrete proper noun with verifiable blog, repo and weight pages — not a rumor or versioned spam.

Usage

Get started

  1. Download weights from Hugging Face or ModelScope (Apache-2.0).
  2. Follow the GitHub Quick Start / Inference Backends: Transformers for experiments; prefer official SGLang for serving/eval; try llama.cpp for local self-speculative decoding.
  3. Apply for hosted preview access at diffuspace.ai (chat entry points per the site).

Tips

  • Treat benchmark and throughput figures as vendor claims until you reproduce them on your tasks and hardware.
  • The model builds on the Qwen3.6-35B-A3B line; block diffusion differs from stock autoregressive Qwen tooling.
  • The "decision model" is a later release; this Preview is the language-model weights.