David Wang
80 posts
I keep getting sick recently. this sucks. I wish my immune system wasn't so terrible. maybe I should replace it with an AI
English
David Wang retweetledi
DFlash⚡ meets OpenClaw🦞 = FlashClaw
Same Claw. >4X faster or cheaper.
DFlash support for Qwen3.5 is live — outperforming native MTP by up to 2.3X.
More to come! 🔥
English
@Dorialexander @AmpCode we have instructions to set up opencode with our glm5 endpoint here: modal.com/blog/try-glm-5 :)
English
So I want with @AmpCode and, unfortunately… (and no, most of my students did not create accounts in time). What is the best recommended free(mium) alternative now?
Alexander Doria@Dorialexander
What is the recommended free/freemium alternative to claude code? GLM on OpenCode? Codex works with free gpt? (for my students, so that they can start without subscription).
English
David Wang retweetledi
@lisperati Just write assembly code. That has always been allowed.
English
GIF
Charles 🎉 Frye@charles_irl
The GLM models by @Zai_org have been a gamechanger for me. I was reluctant to embrace coding agents before I could run the models myself. Now, with GLM-5, I have a top-quality self-hosted intelligence endpoint tightly integrated into my engineering work. github.com/modal-projects…
ZXX
David Wang retweetledi
Check out our paper and code!
Follow our Hugging Face — more draft models for popular LLMs coming soon 🚀
HF: huggingface.co/collections/z-…
Code: github.com/z-lab/dflash
Paper: arxiv.org/abs/2602.06036
Zhijian Liu@zhijianliu_
The paper is now available: huggingface.co/papers/2602.06… More updates coming soon!
English
David Wang retweetledi
The paper is now available: huggingface.co/papers/2602.06…
More updates coming soon!
Zhijian Liu@zhijianliu_
Holiday cooking finally ready to serve! 🥳 Introducing DFlash — speculative decoding with block diffusion. 🚀 6.2× lossless speedup on Qwen3-8B ⚡ 2.5× faster than EAGLE-3 Diffusion vs AR doesn’t have to be a fight. At today’s stage: • dLLMs = fast, highly parallel, but lossy • AR LLMs = accurate, sequential, but slow DFlash = diffusion drafts, AR verifies.
English
Anyone has a job that needs strong OS/System research and engineering skills?
My postdoc Jongyul Kim is looking for a research oriented job. He does great work on Storage and Memory Systems. You can find his work at: yulistic.github.io
His CV can be found at: yulistic.github.io/files/CV_jongy…
English
just tripled my money (throughput per dollar) by predicting multiple tokens at once. amazing alpha (acceptance rate) here
Charles 🎉 Frye@charles_irl
I went into AI, not crypto, but I still ended up speculating on tokens.
English
docker pull modalresearch/sglang:v0.5.7-fa4-preview
Erik Bernhardsson@bernhardsson
It’s ironic that Blackwells have been out since 2024 but people still prefer Hoppers because the kernels aren’t Blackwell-optimized yet, and now the Hopper prices are going up.
English
David Wang retweetledi
David Wang retweetledi
⚡ Speed of flash. Just 2 days after launch, DFlash is already running in SGLang (@sgl_project).
With serving-engine support, we can now unlock speedup with higher concurrency, and we’ve quickly worked on a new demo based on it.
We'll be cooking up more and better draft models over the next few weeks.🔥 Stay tuned!
Akshat Bubna@akshat_b
Two days since DFlash was released, and @_dcw02 (on @modal research) already shipped support for it in SGLang. Why are we so excited about this? Diffusion speculators let us get *way* higher tok/s than auto-regressive models. E.g. we're seeing a 4.73x boost with H200s + FA3 already — with still more improvements to come! Reach out to us if we can help you get this in prod today, and huge thanks to @zhijianliu_ and team for coming up with this technique.
English
Two days since DFlash was released, and @_dcw02 (on @modal research) already shipped support for it in SGLang.
Why are we so excited about this?
Diffusion speculators let us get *way* higher tok/s than auto-regressive models. E.g. we're seeing a 4.73x boost with H200s + FA3 already — with still more improvements to come!
Reach out to us if we can help you get this in prod today, and huge thanks to @zhijianliu_ and team for coming up with this technique.
English
@zhijianliu_ and @jianchen1799 are really amazing, they have the Mandate of Heaven
English
friendship ended with EAGLE3 now DFlash is my best friend
Akshat Bubna@akshat_b
Two days since DFlash was released, and @_dcw02 (on @modal research) already shipped support for it in SGLang. Why are we so excited about this? Diffusion speculators let us get *way* higher tok/s than auto-regressive models. E.g. we're seeing a 4.73x boost with H200s + FA3 already — with still more improvements to come! Reach out to us if we can help you get this in prod today, and huge thanks to @zhijianliu_ and team for coming up with this technique.
English
David Wang retweetledi
We want to make B200s competitive with H100s with open source inference engines.
You're welcome, Jensen!
* github.com/Dao-AILab/flas…
* github.com/Dao-AILab/flas…
* github.com/Dao-AILab/flas…
English
@natolambert @humansand human sand would be a metal name for a chip making company
English
I'm a fan of a lot of folks going to @humansand
unfortunately it'll be referred to as human-sand and not humans-and. Kind of ominous.
Can't unsee it. It's your viral marketing strategy.
English