← News

ResearchSources: arXiv · 19 Nov 2026

Clover-35B-A3B-Preview pairs diffusion generation with recurrent memory for agent tasks

The technical report describes a 35-billion-parameter mixture-of-experts diffusion language model whose authors say it leads seven agent benchmarks, including SWE-bench Verified.

In plain words

A research team has previewed a large language model that generates text through a diffusion-style process rather than one token at a time. The model, Clover, keeps a running memory of text it has already produced while revising blocks of text, and its authors report strong results on coding-agent benchmarks. All figures come from the team's own report and have not been independently verified.

What was announced

Clover-35B-A3B-Preview is a mixture-of-experts diffusion language model with 35 billion total parameters, of which 3 billion are activated per step, and it supports thinking-mode generation. The architectural change adapts Gated Delta Networks to block diffusion: recurrent memory of previously generated context is read out and competes with tokens in the current block in a shared softmax, alongside learnable block-size calibration and a zero-initialized right-context convolution branch. The authors report that Clover leads all seven agent metrics they evaluated, scoring 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1.

On internal agent workloads, the team reports a generation speed of 680 tokens per second on a single NVIDIA GB300 GPU at BF16 precision with batch size one, and the evaluation is organized around model quality, drafting and verification costs, and deployment under constrained memory. The report also describes a decision-model variant that answers related questions in a single denoising pass; on an internal set of 500 groups, joint prediction reportedly raised question accuracy from 90.37% to 93.88% and all-correct group accuracy from 65.4% to 77.0%. These are the authors' own measurements on internal workloads, not independent evaluations.

What it does not show

  • Evidence is limited to the report's abstract, so architectural details, benchmark methodology and comparisons are not available for verification.
  • All benchmark scores, speed figures and accuracy gains are self-reported by the authors on internal workloads; no independent confirmation is provided.
  • The decision-model comparison uses an internal set of 500 groups, and the report excerpt does not describe how those groups were selected.

For security and IT teams

Scope
The reported 680 tokens/s figure applies to one NVIDIA GB300 GPU at BF16 precision with batch size one on internal agent workloads, per the authors.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.