Meet Clover-35B-A3B-Preview 🍀 the first open-weight production diffusion language model (dLLM) with 35B total parameters capable of 680 tps on a single card at BF16. Clover exhibits strong performance against both dLLM and AR models and is available today at @huggingface @ModelScope2022 and at our website

Oct 28, 2026 · 3:42 AM UTC

52
81
657
78,153
DiffuSpace is working with partners to explore cutting edge use cases such as drone logistics and navigation. If that sounds interesting, our DMs are open
2
1,166
Sort replies: Relevant Recent Liked
Replying to @DiffuSpace_Labs @huggingface
Decentish performance compared to a 6 months old model
1
10
938
Replying to @DiffuSpace_Labs @huggingface

ALT Dj Khaled Keys GIF by Music Choice
6
1,018
Replying to @DiffuSpace_Labs @huggingface
680tps? I can't wait to try it out.
160
Replying to @DiffuSpace_Labs @huggingface
A fidsy open weight diffusion models with 680tps. This is wild.
1
573
Replying to @DiffuSpace_Labs @huggingface
BIG THANKS!!!!! Great release!
75
Replying to @DiffuSpace_Labs @huggingface
So fast ！！！
658
Replying to @DiffuSpace_Labs @huggingface
35b-a3bクラスのモデルとして使うのが楽しみです！
187
Replying to @DiffuSpace_Labs @huggingface
Looks really promising but why not compare with mercury 3?
1
464
Replying to @DiffuSpace_Labs @huggingface
Amazing! Tried it and results are dope
191
Replying to @DiffuSpace_Labs @huggingface
@grok no common inference engine supports this as of now right?
1
25
Replying to @DiffuSpace_Labs @huggingface
how does this compare against Mercury 3
91
Replying to @DiffuSpace_Labs @huggingface
I love it please can we get gguf file too it would be best i have 16gb vram i would love to run this model
437
Replying to @DiffuSpace_Labs @huggingface
680 tps on one card is the number that matters. diffusion LLMs could make local AI feel instant instead of "ok for a local model". curious who gets it running on Apple Silicon first, that's where all 493 on-device apps i track on siliconshelf.app live.

SiliconShelf – on-device AI apps for iPhone, iPad and Mac

Searchable directory of iPhone, iPad and Mac apps that run AI models on-device. Filter by platform, price and open source.

siliconshelf.app
4
Replying to @DiffuSpace_Labs @huggingface
Amazing! Can be a very good base for next version of oui-1 models.
1
528
Replying to @DiffuSpace_Labs @huggingface
Letting the GDN history readout compete in the same softmax as the in-block tokens is a genuinely neat trick. But 680 tok/s is bs=1 on internal agent data. Honestly skeptical the edge holds at serving batch sizes, where AR + spec decoding amortizes well. Any bs>1 numbers?
223
Replying to @DiffuSpace_Labs @huggingface
Very glad to see open source with these models! Any chance of a MLX version for Apple Silicon?
1
53
Replying to @DiffuSpace_Labs @huggingface
huggingface.co/Diffuspace/Cl… making your life easier

Diffuspace/Clover-35B-A3B-Preview · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
142
Replying to @DiffuSpace_Labs @huggingface
congrats!
3
631
Replying to @DiffuSpace_Labs @huggingface
I always pictured LLMs writing one word at a time, so filling whole blocks in parallel is a brand-new mental model for me. Does quality hold up when you push for more speed?
50
Replying to @DiffuSpace_Labs @huggingface
If Clover's speed holds up in real workloads, it could make diffusion models a practical option for latency-sensitive products, not just an interesting alternative to AR models.
342
Replying to @DiffuSpace_Labs @huggingface
Most models write left-to-right, one token at a time. Clover fills blocks in parallel — 680 tps on one card. That’s not “faster autocomplete.” That’s a different shape of thinking. Want a playable Diffusion Fill specimen that lights up by confidence?
277
Replying to @DiffuSpace_Labs @huggingface
Now mix it with MemT
107
Replying to @DiffuSpace_Labs @huggingface
When are you going to release the backend components?
37
Replying to @DiffuSpace_Labs @huggingface
Congrats on shipping open weights. Since Clover is derived from Qwen3.6-35B-A3B, how much of the agentic gain comes from diffusion decoding itself vs. the extra post-training?
227
Replying to @DiffuSpace_Labs @huggingface
does the 73.6 on SWE-bench Verified come from the same decoding settings that hit 680 tps, or are those separate configs?
160
Replying to @DiffuSpace_Labs @huggingface
Congrats!
402
Replying to @DiffuSpace_Labs @huggingface
Waiting for the hugging face link
350
Replying to @DiffuSpace_Labs @huggingface
@grok 帮我看看这几个模型的架构，他们是 block diffusion 么？还是做的 dflash 那类？他们怎么解决的 gdn 的问题？解码是 thresholddecoding ？
1
387
Replying to @DiffuSpace_Labs @huggingface
Huge milestone 35B open weight dLLM hitting 680 tps on a single card is insane congrats
401
Replying to @DiffuSpace_Labs @huggingface
@grok 他们是怎么解决gdn的问题的
281
Load more