Skip to content
Modelwire
Subscribe

Clover-35B achieves agent benchmarks with 3B activated parameters per token

Source published ·Modelwire updated

Original coverage: arXiv cs.CL ↗·How Modelwire adds context

The development

Clover-35B-A3B-Preview represents a shift toward efficient mixture-of-experts architectures in reasoning-capable language models. The 35B parameter model activates only 3B parameters per token, combining Gated Delta Networks with block-level diffusion to balance memory and bidirectional context. Benchmark results on agent tasks (73.6 SWE-bench, 51.7 Terminal-Bench) and reported 680 tokens/s throughput on single-GPU inference suggest the approach trades parameter scale for architectural efficiency. This matters for practitioners evaluating inference cost versus reasoning capability in production deployments.

Modelwire’s AI-generated summary of coverage from arXiv cs.CL.

Modelwire analysis

Analyst take

Our AI-generated reading of the wider context and the next developments to watch.

Clover-35B achieves its throughput (680 tokens/s single-GPU) by combining gated routing with block-level diffusion, a specific architectural choice not yet common in the open literature. The benchmark pairing (SWE-bench for reasoning, Terminal-Bench for agent tasks) also signals a deliberate positioning toward software engineering and autonomous systems, narrower than general-purpose reasoning.

This lands in the same efficiency race as Alibaba's Qwen4-Max (2.4T parameters, 71B active per token, launched days earlier) and Hugging Face's 180B MoE work from the same day. All three target the same customer problem: frontier capability at lower inference cost. The difference is scale positioning. Qwen4-Max competes on absolute performance at lower price; Clover-35B competes on per-token efficiency within a smaller parameter budget. The prior work on MoE routing (Fuse-All, EchoMoE, CECAR) addressed infrastructure constraints. Clover-35B assumes those are solved and focuses on architectural trade-offs instead.

If Clover-35B's SWE-bench and Terminal-Bench scores hold up when re-evaluated on held-out test splits (not the preview set), the efficiency claim is credible. If the model ships as open-source within 60 days and achieves similar throughput on commodity A100s (not just GB300), it signals serious deployment intent rather than a paper contribution. Otherwise, treat the 680 tokens/s figure as GB300-specific and discount accordingly for practitioners on older hardware.

This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error

Coverage behind this analysis

These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.

  1. ·The Decoder

    Alibaba's Qwen4-Max cuts inference costs with sparse activation design

    Alibaba's Qwen4-Max represents a significant shift in frontier model economics. The 2.4 trillion parameter architecture activates only 71 billion parameters per token, using a hybrid design that substantially reduces inference costs on extended contexts. Performance lands near Kimi K3.1 Pro on industry benchmarks while undercutting established competitors on pricing. The planned open-source release of smaller…

    Read Modelwire coverage →Original source ↗

MentionsClover-35B-A3B-Preview · Gated Delta Networks · SWE-bench Verified · Terminal-Bench 2.1 · NVIDIA GB300 · arXiv

MW

How this coverage is produced

Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Clover-35B-A3B-Preview Technical Report”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.

More to explore

Hugging Face runs 180B-parameter MoE models on 32GB RAM

Hugging Face·

Frozen large models guide smaller decoders through hidden-state bridges

arXiv cs.CL·

TPU kernel optimizes mixture-of-experts inference through fused computation

arXiv cs.LG·
Clover-35B achieves agent benchmarks with 3B activated parameters per token · Modelwire