Clover-35B-A3B Claims 73.6 on SWE-bench Verified. Every Number Is the Vendor's, and the Decision Model Is Not Out.
DiffuSpace released Clover-35B-A3B-Preview on October 30: a diffusion mixture-of-experts model derived from Qwen3.6-35B-A3B, 35 billion total parameters with about 3 billion active, a 128K context, and an Apache license, with weights on Hugging Face. The benchmarks are DiffuSpace's own: SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7, GPQA Diamond 83.8, AA-LCR 72.0, and 680 tokens per second on a GB300. A companion decision model is promised for a later date with an internal score against Jev. There is no Artificial Analysis entry, and the blog says apply for access while the weights are public. We read the Hugging Face blog and model card.
By T.W. Ghost
DiffuSpace, the group descended from HKU's Dream 7B work, released Clover-35B-A3B-Preview today, October 30, 2026: a diffusion mixture-of-experts language model built on Qwen3.6-35B-A3B, with 35 billion total parameters, about 3 billion active, a 128K context, and an Apache license, with weights on Hugging Face. It is the second text diffusion model this month to claim it beats autoregressive peers on speed, after Mercury 3, and like Mercury it arrives with a benchmark table its maker ran.
The headline number is SWE-bench Verified at 73.6. The others: Terminal-Bench 2.1 at 51.7, GPQA Diamond at 83.8, AA-LCR at 72.0, and agent and tool scores against older Qwen3.6 builds. Speed, also DiffuSpace's: 680 tokens per second on a GB300, 515 on an H200 against 365 for Qwen with DFlash2, and 101.8 tokens per second at INT4 on a 16 GB RTX 5070 Ti. Two things are not shipped: the decision model the blog pairs with it, "released at a later date," with an internal Joint Mode score of 95.0% against Jev's 63.6%; and any independent score, since Artificial Analysis has no row. The blog also says "apply for access" on a page whose weights are already public, which is marketing, not a gate. We read the Hugging Face blog and the model card. Here is what held up.
What You Are Downloading
| Clover-35B-A3B-Preview | |
|---|---|
| Released | October 30, 2026 |
| Architecture | Diffusion mixture of experts; 35B total; about 3B active; Qwen3.6-35B-A3B backbone |
| Context | 128K (model card) |
| License | Apache (blog and card); confirm the LICENSE file |
| Weights | Public on Hugging Face |
| Decision model | Not released; "at a later date" |
| Independent score | None (no AA row) |
Three billion active on a 35-billion base is the efficiency story. It is also the Qwen3.6 story, since the backbone is Qwen's. What DiffuSpace adds is the diffusion decoding, and the speed claims are the measurement of that.
"Apply for access" and "weights on Hugging Face" cannot both be the gate. The weights are the fact; the form is the funnel.
The Benchmarks, All of Them DiffuSpace's
| Benchmark (DiffuSpace's blog) | Clover-35B-A3B |
|---|---|
| SWE-bench Verified | 73.6 |
| Terminal-Bench 2.1 | 51.7 |
| GPQA Diamond | 83.8 |
| AA-LCR | 72.0 |
| Speed, GB300 | 680 tokens per second |
| Speed, H200 | 515 tokens per second (Qwen with DFlash2: 365) |
| Speed, INT4 on 16 GB RTX 5070 Ti | 101.8 tokens per second |
| Decision model, Joint Mode (internal set) | 95.0% vs Jev 63.6%; model not released |
Three honest observations:
- The comparators are the vendor's choice. Mostly Qwen3.6-class peers, measured by DiffuSpace, on DiffuSpace's harness. A 73.6 on SWE-bench Verified from a 3B-active model would be remarkable, which is the reason it needs a second source.
- Terminal-Bench 2.1 is the old version, the same note we made for Nanbeige4.3-33B and Recursive Flint this month. It is not the 4.0 figure this site uses for frontier comparisons.
- The decision model is a promise with a number attached. A 95.0% on an internal set against Jev, for a model you cannot download, is a roadmap item. Mercury made a similar companion claim; we treated it the same way.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Oct 30; 35B total, about 3B active; Qwen3.6-35B-A3B backbone; 128K | Verified (model card and blog) |
| Apache license | Verified on blog and card; LICENSE file not confirmed here |
| SWE-bench Verified 73.6; Terminal-Bench 2.1 51.7; GPQA Diamond 83.8; AA-LCR 72.0 | Vendor |
| 680, 515, 101.8 tokens per second | Vendor |
| Decision model Joint Mode 95.0% vs Jev 63.6% | Vendor, internal set, model unreleased |
| Artificial Analysis score | Unverified (none) |
| Any independent re-benchmark | Unverified (none) |
Who Should Use It
Try Clover if you want a fast, small-active-parameter open model under Apache and you can benchmark it on your own repository; the architecture and license are the verified parts.
Skip the SWE-bench line until someone other than DiffuSpace runs it, and skip planning around the decision model until it exists.
The frontier is elsewhere: the diffusion models this month have speed tables and no independent intelligence scores, and the open models with independent scores are on a board that was just rebuilt. Our model comparison has the current standings side by side.
Sources
- Hugging Face blog: Clover-35B-A3B-Preview
- Hugging Face: Diffuspace/Clover-35B-A3B-Preview
- GitHub: DiffuSpace/Clover-35B-A3B-Preview
*Not sure whether a diffusion model's speed is worth an unscored intelligence level? Take the free 2-minute quiz and get matched. Then read about the other text diffusion launch this month and the gap between its vendor speed and the measured one: Mercury 3.*