DiffuSpace Clover is the open diffusion language model that has quietly become one of the most talked-about coding releases of late October 2026. Instead of writing text one word at a time like almost every chatbot you have used, DiffuSpace Clover writes in whole blocks of 32 tokens at once, filling in the pieces it is most sure about first and revisiting the rest. The result, according to its maker, is a free, downloadable model that fixes real software bugs faster than its own base model and generates code at hundreds of tokens per second.
That is a bold promise, and it arrives with an important asterisk: DiffuSpace Clover is a preview release, and its headline numbers come from the company itself, not from independent testing. In this guide, we explain what Clover actually is, how block diffusion works in plain English, what the benchmarks really say, what hardware you need to run it, and — just as importantly — what you cannot do with it yet. By the end, you will know whether Clover deserves a place in your coding workflow or whether it is one to watch from the sidelines for now.
Table of Contents
- What Is Clover?
- How Clover Writes in Blocks, Not Words
- Clover Specs at a Glance
- Clover Benchmarks: What the Numbers Say
- Clover Speed: 515 Tokens per Second?
- How to Run Clover Today
- Clover vs Qwen, Mercury 3, and Regular Coding Models
- Who Should Use Clover — and Who Should Wait
- Clover Limitations and Honest Caveats
- Clover Pricing and Licence
- Clover FAQs
- The Verdict on Clover
What Is DiffuSpace Clover?
DiffuSpace Clover, formally named Clover-35B-A3B-Preview, is the first model released by DiffuSpace, a Shenzhen-based company focused entirely on diffusion language models. It is a text-only, open-weight model built for coding and agentic work — the kind of long, multi-step tasks where an AI assistant reads a codebase, plans a fix, edits files, and checks its own work.

The name tells you most of the story. The “35B” refers to 35 billion total parameters, while “A3B” means only about 3 billion of those parameters are active for any given token. That is the mixture-of-experts trick: the model is split into specialist sub-networks, and each token consults only a few of them. You get much of the capability of a large model at a fraction of the per-token compute cost.
DiffuSpace Clover and the Team Behind It
DiffuSpace is not a newcomer that appeared overnight. The team previously built Dream 7B, an earlier diffusion model, at the University of Hong Kong. In October 2026, the company announced it had completed two funding rounds totalling several hundred million Chinese yuan, led by Matrix Partners China, Shunwei Capital, and Legend Capital, with participation from investors including Huawei Hubble and Horizon Robotics. The company described it as the largest funding round ever raised globally by a diffusion language model company.
That backing matters because diffusion language models are still a challenger technology. The established giants — OpenAI, Google, Anthropic, and Alibaba — all build traditional autoregressive models. DiffuSpace is betting that a different way of generating text can win on speed and cost, especially for coding agents that burn through enormous numbers of tokens in long sessions.
DiffuSpace Clover is also a deliberate follow-up to a wider trend. Inception’s Mercury 3, covered in our Inception Mercury 3 guide, proved that diffusion models could reach frontier-adjacent speeds. Google has its own text diffusion work, and Amazon has experimented in the same direction. Clover is the first open-weight entry in this race that ordinary developers can download and run themselves.
How DiffuSpace Clover Writes in Blocks, Not Words
Nearly every language model you have used writes from left to right, one token at a time. A token is a small chunk of text — often a word or part of a word. The model picks the most likely next token, appends it, and repeats. It cannot go back and revise an earlier choice, and it cannot look ahead. That is the autoregressive approach, and it works remarkably well, but it has a built-in speed limit: a 500-token answer needs 500 sequential steps.
DiffuSpace Clover breaks that rule. It starts each 32-token block as a row of masked placeholders, then fills in every position it is at least 95% confident about in parallel. Positions it is unsure about stay masked for another pass. The process repeats until the whole block is complete, and then the finished block becomes part of the context for the next one. In effect, the model drafts in parallel and revises as it goes, rather than committing to each word in order.
DiffuSpace Clover LowConfidence Decoding Explained
The mechanism has a name: LowConfidence decoding. At each pass, the model scores its confidence for every masked position. Confident positions get filled immediately; uncertain ones wait. If no position clears the confidence threshold, the single most confident position is filled anyway, so the process always makes progress. Two generation controls ship with the model: a repetition penalty to stop very long answers from looping, and gradual biases that nudge the model toward ending its response and its thinking phase cleanly.
Why does this matter for coding? Because code is full of places where the model already knows the answer with near-certainty — closing brackets, boilerplate imports, repeated patterns — and a few places where real thought is needed. A block-diffusion model can blast through the easy parts in parallel and spend its passes on the hard parts. It can also, in principle, fix an inconsistency inside a block before moving on, something a left-to-right model can never do.

DiffuSpace Clover Architecture Under the Hood
Under the hood, DiffuSpace Clover keeps the hybrid backbone of its base model, Alibaba’s Qwen3.6-35B-A3B. The model card describes 40 layers mixing Gated DeltaNet (GDN) attention with full attention, 256 routed experts with top-8 routing per token, and one shared expert per layer. DiffuSpace adds block diffusion inference, bidirectional GDN parameters, and its own generation controls on top of that foundation.
One detail worth noting: DiffuSpace Clover is text-only. It cannot see images, read screenshots, or process audio. If your workflow needs vision — reviewing a UI mockup, reading a diagram — this is not the model for it. Its context window is 131,072 tokens shared between prompt and response, which is generous for coding sessions but not class-leading in 2026.
DiffuSpace Clover Specs at a Glance
| Specification | DiffuSpace Clover |
|---|---|
| Full name | Clover-35B-A3B-Preview |
| Maker | DiffuSpace (Shenzhen) |
| Release date | Announced October 28, 2026; coverage landed October 29, 2026 |
| Total parameters | 35 billion |
| Active parameters per token | About 3 billion |
| Architecture | Mixture-of-experts with block diffusion, 40 layers, 256 routed experts, top-8 routing |
| Context window | 131,072 tokens (prompt + response combined) |
| Block size | 32 tokens by default, confidence threshold 0.95 |
| Modalities | Text only |
| Base model | Alibaba Qwen3.6-35B-A3B |
| Weights size (BF16) | About 69 GB |
| Licence | Apache 2.0 (upstream Alibaba assets included with attribution) |
| Where to get it | Hugging Face (Diffuspace/Clover-35B-A3B-Preview), GitHub, ModelScope mirror |
| Status | Preview — benchmarks are company-reported, not independently verified |
DiffuSpace Clover Benchmarks: What the Numbers Say
Every number in this section comes from DiffuSpace’s own launch materials, tested with the model’s thinking mode on. No independent laboratory has reproduced them yet, and the company has not published the full evaluation scaffolding — the agent harness, tool setup, attempt counts, and runtime details that would let outsiders check the scores. Treat the figures as a claim to be tested, not as settled fact. That said, the pattern of the claims is interesting, because DiffuSpace Clover beats its own base model exactly where a coding model should, and trails it where diffusion might be expected to struggle.
The headline claim is 73.6 on SWE-bench Verified, against 65.8 for the Qwen3.6-35B-A3B base model. SWE-bench Verified is one of the most respected coding tests: 500 real software issues, reviewed by humans, drawn from real repositories. The model is given the code and the issue description and must produce a patch that actually fixes the problem. If 73.6 is a plain percentage of those 500 issues, it means roughly 368 issues resolved — a strong result for a model of this size, and one that would put DiffuSpace Clover in serious company.
On agent-style coding, DiffuSpace Clover reports 51.7 on Terminal-Bench 2.1, up from 43.1 for the base model. Terminal-Bench tests whether an agent can complete real tasks in a terminal environment, not just write a snippet in isolation. That jump matters because agentic coding — long sessions with tools, files, and feedback — is precisely the workload DiffuSpace is targeting.
Two more company-reported scores round out the picture: 65.3 on tau-cubed (tau-bench), a test of tool use in realistic scenarios, and 67.9 on BFCL-v4, which measures function calling — the ability to invoke the right tool with the right arguments. Both are the kind of unglamorous, practical capabilities that decide whether a coding agent is actually useful or merely impressive in demos.
General reasoning looks steadier rather than spectacular: 83.8 on GPQA Diamond, a hard science question set, versus 83.3 for the base model. And the honest weak spots are visible too. DiffuSpace Clover trails its base model on AIME 2026 mathematics (83.3 versus 90.4) and on HLE text-only (12.8 versus 21.2). A company that publishes its losses alongside its wins deserves some credit — and it tells you this model is specialised for coding and tool use, not a general-purpose reasoning champion.

DiffuSpace Clover Speed: 515 Tokens per Second?
Speed is the whole point of diffusion, so the speed claims deserve the closest look. DiffuSpace reports 515 tokens per second on a single NVIDIA H200 GPU, and a separate industry digest reports 680 tokens per second on a single GB300 chip. For comparison, a typical autoregressive model of this size on similar hardware generates far more slowly, because each token must wait for the previous one.
There is also a smaller-hardware claim worth mentioning with caution: DiffuSpace has been reported as claiming 101.8 tokens per second on MATH-500 with INT4 quantization on an RTX 5070 Ti with 16 GB of memory, against 48.5 tokens per second for similarly sized MiniCPM and Qwen models. Independent writers have already questioned that figure, noting that a 35-billion-parameter model at 4-bit precision is larger than 16 GB, and that DiffuSpace has not said which runtime, quantization recipe, or offloading setup it used — nor released those INT4 weights. Treat it as a preview of what might come, not as something you can reproduce today.
The fair way to think about Clover speed is this: the architecture genuinely removes the one-token-at-a-time bottleneck, and the H200 figure is at least plausible for a 3-billion-active-parameter model generating in parallel blocks. But until an independent tester runs the same prompts on the same hardware with a published harness, the exact numbers remain marketing. If you are choosing infrastructure on the basis of tokens per second, wait for that independent run.
How to Run DiffuSpace Clover Today
Here is where the excitement meets reality. Clover is genuinely downloadable today — the weights and code sit on Hugging Face and GitHub, and anyone can pull them down. But running them is, for now, a job for serious hardware and patient developers, not a one-click install.
The only officially working path at launch is the Hugging Face Transformers route, which needs Transformers 5.19.0, PyTorch 2.10 or later, CUDA, and an NVIDIA GPU with about 69 GB of memory for the BF16 weights. The official release is 26 safetensors shards totalling roughly 69 GB, with a ModelScope mirror for users who cannot reach Hugging Face. In practice, that means a data-center-class GPU such as an H100 or H200, or a multi-GPU setup — not a laptop, and not a typical gaming PC.
The model card recommends the company’s SGLang implementation for serving and evaluation, warning that small numerical differences between backends can produce different text. But as of late October 2026, independent checks found the SGLang support still sitting as an unmerged pull request, and the llama.cpp backend code listed as “will be added” rather than shipped. The official GitHub repository contains a README, a benchmark chart, and an empty folder where further code is promised. There is no GGUF release, no Ollama model, and no LM Studio build. If you want this model family on your own hardware this week, the honest shortcut is to run the base Qwen3.6-35B-A3B model instead, which already works everywhere.
That gap will probably close quickly — open-weight releases usually gain community backends within weeks — but anyone writing a “how to run it” guide in the first days of the release should say so plainly. Clover is downloadable, but it is not yet plug-and-play.
DiffuSpace Clover Setup Checklist
- GPU: NVIDIA GPU with around 69 GB of free VRAM for the full BF16 weights (H100/H200 class, or multiple smaller cards).
- Software: Transformers 5.19.0, PyTorch 2.10+, CUDA, and the official model repository code.
- Download: Hugging Face repository Diffuspace/Clover-35B-A3B-Preview (26 shards, about 69 GB).
- Defaults: Block size 32 tokens, confidence threshold 0.95 — the settings behind the published benchmarks.
- Not yet available: Merged SGLang backend, llama.cpp build, GGUF weights, Ollama or LM Studio packages.

DiffuSpace Clover vs Qwen, Mercury 3, and Regular Coding Models
Against its own base model, Clover is a clear improvement for its target workload: SWE-bench Verified up from 65.8 to 73.6, Terminal-Bench 2.1 up from 43.1 to 51.7, with general science roughly flat. That is the expected signature of a good fine-tune plus a better decoding method: the model gets better at the thing it was tuned for, without becoming a different model entirely.
Against Inception Mercury 3, the comparison is about philosophy as much as numbers. Mercury 3 is a commercial, API-first diffusion model built for raw speed — 1,068 tokens per second claimed — with a 1-million-token context window and a pricing model aimed at enterprises. Clover is the open-weight answer: free to download, runnable on your own hardware, with a much smaller 128K context and a preview label. If you want an API you can call today, Mercury 3 is the practical choice. If you want weights you can inspect, modify, and host yourself, Clover is the only diffusion option in this class.
Against traditional autoregressive coding models, the honest answer is that Clover has not yet been tested head-to-head by independent reviewers. Our best AI coding assistant guide compares the established names — GitHub Copilot, Cursor, Claude Code, and others — on price, agent mode, and best-for-whom. Clover is not a product in that sense; it is a raw model. To use it as a coding assistant, you still need an agent harness, an editor integration, and tooling that the community has not yet built. Its natural comparison set is other open-weight models such as Qwen, DeepSeek, and Mistral — including the newly released open weights covered in our Mistral Large 4 open weights guide — rather than finished coding products.
Who Should Use DiffuSpace Clover — and Who Should Wait
Try Clover now if you are: a researcher studying diffusion language models; a team building coding agents who can afford H100-class hardware and wants to experiment with parallel decoding; an open-weight enthusiast who enjoys testing preview releases and reporting what breaks; or a company evaluating whether diffusion could cut token costs in long agent loops.
Wait a few weeks if you are: a solo developer on a laptop or single consumer GPU; someone who needs an Ollama, LM Studio, or API endpoint today; a team that needs independently verified benchmarks before committing; or anyone whose workflow depends on vision, image input, or a context window beyond 128K tokens. For those users, the base Qwen model, or a finished product from our coding assistant guide, is the better choice right now.
The most exciting audience may be agent builders. Coding agents are the workload where diffusion’s speed advantage compounds: an agent that runs dozens of tool calls and reasoning steps per task pays the token-by-token tax on every step. If Clover delivers even half its claimed speed in real agent loops, it changes the economics of running agents at scale. That is also why the company promises a decision-model variant to follow — a model tuned for the planning and tool-selection side of agentic work, though no date or details are available yet.
DiffuSpace Clover Limitations and Honest Caveats
- Preview status: This is a first release with a “Preview” label. Expect rough edges, missing backends, and benchmark revisions.
- Unverified benchmarks: All headline scores are company-reported. The evaluation harness, attempt counts, and runtime details have not been published in enough detail for independent reproduction.
- Hardware demands: About 69 GB of GPU memory for full-precision weights. No consumer-friendly quantized release exists yet.
- Missing backends: SGLang support unmerged, llama.cpp code promised but not shipped, no GGUF, Ollama, or LM Studio builds at launch.
- Text only: No image, audio, or video input. No multimodal capability at all.
- 128K context: Generous for coding, but far behind million-token frontier models for very large codebases.
- Weaker at maths and broad knowledge: Trails its own base model on AIME 2026 and HLE text-only benchmarks.
- Speed claims are hardware-specific: The 515 tokens-per-second figure is for a single H200 GPU under the company’s own test conditions, not a guarantee on your setup.
DiffuSpace Clover Pricing and Licence
Clover itself costs nothing. The weights are distributed under the Apache 2.0 licence, with upstream Alibaba assets preserved under their own Apache 2.0 terms and attribution notices kept in the repository. That means you can use the model commercially, modify it, and build products on it, subject to the usual Apache licence conditions. There is no subscription, no per-token API price, and no paid tier announced.
The real cost is hardware. Running the full BF16 model needs data-center-class GPU capacity, which rents for real money on cloud providers — typically several dollars per hour for an H100-class instance, depending on region and commitment. Self-hosting only becomes cheaper than an API once you run enough tokens to amortise that hardware. For most individual developers, the cheapest way to experiment with Clover today is to wait for community quantized builds, which should shrink the memory requirement dramatically, or to test it on a rented GPU for a short evaluation run rather than buying hardware outright.
There is no official hosted API for Clover at launch. If you see a third-party site offering “Clover API access,” check carefully what model is actually behind it — early in a release cycle, unofficial wrappers sometimes serve the base model or a different model entirely.
DiffuSpace Clover FAQs
Is DiffuSpace Clover free?
Yes. Clover weights are free to download under the Apache 2.0 licence. You pay only for the hardware or cloud GPU time needed to run them. There is no subscription and no per-token charge from DiffuSpace itself.
Can I run DiffuSpace Clover on my laptop?
Not in its current full-precision form. The BF16 weights need about 69 GB of GPU memory, which no laptop provides. Wait for community quantized builds (GGUF or similar) or rent a cloud GPU for a short test. The base Qwen3.6 model already runs on much smaller hardware if you need something today.
Is DiffuSpace Clover better than GPT or Claude for coding?
We do not know yet. Clover beats its own Qwen base model on company-reported coding benchmarks, but no independent head-to-head against frontier closed models has been published. Its appeal is openness, speed, and cost — not a proven claim to beat the best closed models.
What is a diffusion language model?
A diffusion language model generates text by starting from masked placeholders and filling them in parallel over several passes, rather than writing one token at a time from left to right. Clover writes in 32-token blocks, filling confident positions first and revisiting uncertain ones. The approach can be faster and can revise within a block, but it is newer and less battle-tested than traditional generation.
Does DiffuSpace Clover work with Ollama or LM Studio?
Not yet. At launch there is no Ollama model, no LM Studio build, and no GGUF release. The only working path is the Transformers implementation on an NVIDIA GPU. Community backends usually arrive within weeks of an open-weight release, so check the official repository before assuming this is still true.
Is DiffuSpace Clover open source?
The weights and model code are released under Apache 2.0, which is the standard open-weight arrangement used by most open model releases. The full training data and training code are not published, so “open-weight” is the precise description rather than fully open source in the strictest sense.
The Verdict on DiffuSpace Clover
Clover is the most interesting open-weight release of the week precisely because it is both genuinely new and genuinely unfinished. The idea is sound: writing code in parallel blocks should be faster than writing it one token at a time, and the company’s own numbers suggest the approach also improves coding accuracy over its base model. The execution, so far, is a preview: downloadable weights, a working Transformers path, and a set of benchmarks that still need independent checking, surrounded by backends that are promised rather than shipped.
For researchers and agent builders with serious GPU hardware, Clover is worth testing this week — the weights are free, the licence is permissive, and early testers shape what the community builds next. For everyone else, the smart move is to watch: wait for quantized builds, a merged SGLang backend, and the first independent benchmark reproduction. If those three things land, Clover could become the default open diffusion model for coding agents. If they do not, it will remain a fascinating preview of a technology whose time is still arriving.
Either way, the direction of travel is clear. Diffusion is no longer a research curiosity confined to image generation. With Mercury 3 in the commercial world and Clover in the open world, the race to make language models write in parallel has properly begun — and coding agents, with their long, expensive token loops, are where that race will be won or lost.
Sources: DiffuSpace model card on Hugging Face (Diffuspace/Clover-35B-A3B-Preview); AlexTech coverage of the Clover release; Webdiga’s independent runtime check of Clover-35B-A3B; Daily Inference’s analysis of the SWE-bench claim; Hotfix Daily issue 7; RecodeX Pro on DiffuSpace’s funding rounds. All benchmark and speed figures are DiffuSpace-reported unless stated otherwise.
