DiffuSpace

Building the next foundation model paradigm beyond autoregression.

[DiffuSpace @DiffuSpace_Labs]

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

[DiffuSpace @DiffuSpace_Labs]

Love seeing more experiments with Clover!

Excited to see the community pushing its capabilities further. 😄

[DINQ @dinq_me]

Unlike standard autoregressive (AR) decoding, which generates tokens sequentially, Clover uses block diffusion, compressing historical context into Gated DeltaNet states while applying bidirectional attention for parallel denoising.

With 35B total parameters and ~3B active per token, Clover enables faster multi-perspective sampling. In official internal H200 Agent benchmarks, it achieved 515 tokens/s, compared with 365 tokens/s for Qwen + DFlash2.

For candidate analysis, resumes are structured into controlled evidence sets, with career outcomes limited to predefined categories:

p(category | resume, observation level)

Bidirectional interactions help capture relationships between evidence. Repeated sampling evaluates stability, while future career trajectories will be used to assess predictive accuracy.

For local deployment, Cortiq's 4-bit GPTQ model is ~19.36 GB, with 24 GB+ VRAM recommended, including RTX 3090-class GPUs. Additional memory is needed for context and inference states.

In Cortiq 0.8.41 RTX 5090/Vulkan tests, a denoising threshold of 0.8 delivered 124–150 tokens/s in code generation, depending on content and decoding settings.

[DiffuSpace @DiffuSpace_Labs]

Clover 又解锁了一个有意思的应用场景！🍀

感谢@dinq_me 团队的探索和实践，欢迎大家前去体验一下 ，看看 AI 如何让人才分析变得更智能、更高效！🚀

[DINQ @dinq_me]

DINQ × Clover｜让候选人分析千人千面

我们用 Clover＋Jev，在 DINQ 做了一套候选人分析模型。基础数据来自数百万份履历及公开资料：学术个人主页、LinkedIn、Dribbble、Behance，以及 https://t.co/NUoEFxcLrw 聚合的社交信息。清洗后微调 Clover-35B-A3B-Preview，由它生成多视角分析，再由 Jev 完成预设发展类别的决策。
@DiffuSpace_Labs

[DiffuSpace @DiffuSpace_Labs]

From resumes to talent analysis and decision-making — love seeing Clover in action! 🍀

Huge thanks to the @dinq_me team for bringing this use case to life. Can’t wait to see what’s next! 🚀

[DINQ @dinq_me]

DINQ × Clover | Personalized Candidate Analysis for Every Talent Profile

We built a candidate analysis model at DINQ using Clover + Jev. The underlying dataset consists of millions of resumes and publicly available information, including academic personal websites, LinkedIn, Dribbble, Behance, and social profiles aggregated through https://t.co/NUoEFxcLrw.

After data cleaning, we fine-tuned Clover-35B-A3B-Preview to generate multi-perspective candidate analyses, while Jev makes the final decisions across predefined career development categories.

@DiffuSpace_Labs

[DiffuSpace @DiffuSpace_Labs]

@samuel_ys92 Love what you’re building!

[DiffuSpace @DiffuSpace_Labs]

@samuel_ys92 Just the beginning! 🚀

[DiffuSpace @DiffuSpace_Labs]

@ivanfioravanti best time of the year, not too hot and it's not raining.

[DiffuSpace @DiffuSpace_Labs]

Building with @SigmaZ_official 🤝

Our interactive education demo is now live.

Try it here: https://t.co/cejVq5ts3B

More to come as we continue optimizing together!

[SigmaZ AI Lab @SigmaZ_official]

https://t.co/U4oZUN0Fxa

[DiffuSpace @DiffuSpace_Labs]

@tonypythoneer GGUF: https://t.co/4MMSkg9Jpw

[DiffuSpace @DiffuSpace_Labs]

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

[DiffuSpace @DiffuSpace_Labs]

@good2thinkwith I have a mac as well so will try to push the team to get it done

[DiffuSpace @DiffuSpace_Labs]

@harden_369741 what's your setup? Our GGUF version was just released

[DiffuSpace @DiffuSpace_Labs]

@catmanyau that's our goal, fast + reasonable intelligence

[DiffuSpace @DiffuSpace_Labs]

@ddalcu @huggingface would love to make it run on a reasonably powerful mac for local llm, the tps makes it actually useful

[DiffuSpace @DiffuSpace_Labs]

@PatrickBeyer1 @huggingface @grok what inference engine would you like?

[DiffuSpace @DiffuSpace_Labs]

@paleocybernetic @EmircanErkulX @huggingface and timing of when benchmarks are prepared ahead of release, we'll include it in the future

[DiffuSpace @DiffuSpace_Labs]

@yxxshly @S_Man107 @huggingface yeah, at 35B, it's not frontier intelligence, but the 680 TPS make agents respond much faster and you can feel the difference for sure

[DiffuSpace @DiffuSpace_Labs]

@mattapperson @huggingface just timing of things, we'll use later benchmark for future models

[DiffuSpace @DiffuSpace_Labs]

@Greg_GL_87 @huggingface Working on trying to get it on something 16gb with reasonable intelligence

[DiffuSpace @DiffuSpace_Labs]

@math_castello @ManiciCan @huggingface @UnslothAI it's available now!

https://t.co/4MMSkg9Jpw

[DiffuSpace @DiffuSpace_Labs]

@SuperJuicyAncap cooking up something bigger for sure

[DiffuSpace @DiffuSpace_Labs]

GUFF version released, saw someone getting 100 tps on a 3090 🥰

https://t.co/4MMSkg9Jpw

[DiffuSpace @DiffuSpace_Labs]

@DavidGuabb proof not written by ai 🤣

[DiffuSpace @DiffuSpace_Labs]

@DavidGuabb fat fingered, haha