What happened
DiffuSpace announced Clover-35B-A3B-Preview, the first release in its Clover family. The company describes it as a mixture-of-experts diffusion language model with 35 billion total parameters and 3 billion activated parameters, plus a thinking mode. The work builds on Dream 7B, released by the University of Hong Kong NLP group in April 2025; DiffuSpace says team members began the new project from that starting point.
Unlike a conventional autoregressive model that emits tokens sequentially, Clover uses block diffusion. DiffuSpace says its redesigned bidirectional Gated DeltaNet architecture compresses generated history into a recurrent state, while tokens inside the current block interact through bidirectional softmax attention. A separate historical-memory readout participates in the same softmax selection, and the model also uses learnable block-size calibration and a zero-initialized convolution branch for right-side context. The stated objective is to combine compact long-range memory with block-level parallel denoising.
The weights and related code are open source under an Apache license. Developers can obtain them through Hugging Face, GitHub, and ModelScope, and DiffuSpace says the model supports inference through SGLang and llama.cpp. The company also offers an application process for an online trial. This is therefore a deployable model release rather than only a benchmark paper or hosted demonstration.
DiffuSpace reports strong results in coding and tool-use evaluations: 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1, and 65.3 on tau3 Overall. It also reports 83.8 on GPQA Diamond, 95.8 on ZebraLogic, and 72.0 on AA-LCR. In a vendor test on one NVIDIA GB300 GPU, using BF16 at batch size 1 and internal agent-scenario data, it reports 680 tokens per second. On one H200, it reports 515 tokens per second in an agent task, compared with 365 tokens per second for its cited DFlash2-accelerated Qwen comparison.
The announcement includes a local-deployment claim as well: Clover was quantized to INT4 and run on an NVIDIA RTX 5070 Ti with 16GB of VRAM, where DiffuSpace reports 101.8 tokens per second on MATH-500. It also previews a separate decision model, not yet released, intended to answer up to eight related questions together in a joint mode. DiffuSpace says it is developing an agent harness and desktop programming client, but those tools remain in development.
What the archive adds
This is a new event in the wider thread of open-weight MoE releases for coding and agent work, rather than an update to an earlier Clover note. Recent related releases include MiniMax-M3.1-Flash and Step 5 Preview. Clover’s distinguishing claim is its diffusion-generation architecture and associated throughput figures.
Trend signal
No supplied pattern or archived assumption genuinely fits this release closely enough to support a broader pattern claim.
Connections to other events
The earlier open-weight MoE releases from MiniMax and StepFun show a continuing supply of models that teams can run and adapt themselves. Clover joins that set, but its comparison is framed around diffusion generation versus similarly sized autoregressive models, not simply scale, context length, or license.
Why it matters
For a team building coding agents, output latency is not merely a user-interface metric. Each generation can sit inside a loop of planning, tool execution, browser interaction, observation, and correction. A model that is genuinely faster at the same task quality can allow more attempts inside a fixed time budget or make an interactive agent feel less stalled.
Clover offers a practical route to test that premise because its weights are available and the announcement names SGLang and llama.cpp as supported paths. The reported active-parameter count also makes it relevant to teams considering MoE deployment economics. But the useful comparison is not a headline token-per-second number: it is completed, verified tasks under the team’s own tools, hardware, prompts, and latency constraints.
What happens next
- Independent users will test whether Clover’s throughput advantage survives real coding-agent harnesses and tool waits. (3 months, 70%). Watch for: reproducible SGLang or llama.cpp benchmarks using public agent tasks.
- DiffuSpace will release or further detail its separate decision model and agent tooling. (6 months, 55%). Watch for: downloadable artifacts, documentation, or public access rather than a launch claim alone.
Limits and uncertainty
The central performance, speed, and local-deployment figures in the announcement are vendor-reported. The source describes selected benchmark comparisons and testing conditions, but it does not provide independent replication, a full cost comparison, or evidence that results transfer to every coding-agent harness. Its GB300 throughput figure uses internal agent-scenario data, and the H200 comparison is against a specifically accelerated Qwen configuration. The post also does not establish model reliability over long autonomous jobs, safety behavior, production support commitments, or the operational cost of serving it at scale. The planned decision model, harness, and desktop client are not available products in this announcement.
Skill to build now
Create a small evaluation harness that records task completion, wall-clock time, tool-call failures, and human review minutes for a fixed set of coding tasks. Run an autoregressive baseline and Clover with the same tools, prompts, step limits, and hardware settings. This gives the team an answer to the relevant question—whether a faster generation method improves verified work—rather than an answer limited to tokens per second.
Reader takeaway
Clover is an open, deployable 35B/3B-active MoE diffusion model positioned as a speed-oriented option for coding agents. Its reported benchmarks and inference numbers are promising but remain claims from the release source; the next meaningful test is a controlled run in a real agent workflow.
Open question
Can Clover’s reported speed advantage improve verified end-to-end agent completion once browser delays, tool execution, retries, and human review are included?