Higgs Audio v3 TTS is here. Built for voice AI that speaks, not just reads: • 100 languages with single-digit WER/CER • inline control over emotion, style, prosody, and sound effects • API, Workspace, and open weights • Blog 👉 boson.ai/blog/higgs-audio-v3… Watch the demo 👇

gitlab is way worse than github :((

we have become weaker as a species

🚀 The wait is over! Today at #MLSys, we'll give a talk to reveal the final results and present the awards for the FlashInfer AI GPU Competition! 🏆 I'll also introduce FlashInfer-Bench: an agent-oriented Benchmark Engine designed for production kernels. Join us from 11:00 AM - 1:00 PM PT to see who takes the crown and learn more. Everyone is welcome to attend—see you there! ✨ 🌐 Competition & Results: mlsys26.flashinfer.ai 💻 FlashInfer-Bench Benchmark Engine: github.com/flashinfer-ai/fla… #FlashInfer #MLSys26 #AI #GPU

A thread about the history and internal implementation details of activation checkpointing APIs in PyTorch. 🧵

good stuff 🦀

once-in-a-lifetime opportunity to join Yuchen’s team!

#MLSys2026 is happening in two weeks! Our AI Infra team at @perplexity_ai is throwing a happy hour event at Bellevue on May 19. Come chat with us about inference, post-training, RL, kernels, GPUs, RDMA, agents, anything... luma.com/6knju9tp

Introducing XGrammar-2: structured generation for complex agent harnesses. Strict tool-calling formats. Built-in DeepSeek-V4 and Qwen-3.6 support. Up to 80x speedup over XGrammar. Ready-to-use integrations with vLLM, SGLang, TensorRT-LLM, and more! ⚡ From Claude Code to OpenClaw, agents are defining more complex harnesses. XGrammar-2 ensures LLMs always interact with them in the right way. Built in collaboration with DeepSeek, Databricks, and leading frontier AI labs to bring XGrammar-2 into latest models and products. 🧩 Structural Tag: one unified abstraction to describe any format your agent needs 🚀 Scales to 500 strictly typed tools for complex agent harnesses 🌐 Native APIs in Python, C , Rust, and JS, running everywhere from cloud to edge 🛠️ Integrated with vLLM, SGLang, TensorRT-LLM, and more Excited to see what agent builders create with it! Blog: blog.mlc.ai/2026/05/04/xgram… GitHub: github.com/mlc-ai/xgrammar

Excited to share that I’ll be presenting SkyWalker at #EuroSys26 in Edinburgh tomorrow!🚀 We asks: Can we reduce the cost of multi-region LLM serving by cross-region offloading, without losing the benefits of KV-cache locality? Talk: April 29, afternoon track A, ~16:20-16:40📍

TIL: python’s set doesn’t maintain insertion order unlike dict which does 😅

The DeepSeek proposals section is always fun lol

TileLang is DeepSeek’s kernel maxxxing DSL on both Nvidia and Huawei’s Ascend cards

In this paper is presented Event Tensor, an abstraction designed to simplify the compilation and execution of dynamic megakernels, providing first-class support for both shape and data-dependent dynamism. arxiv.org/pdf/2604.13327

Why subscribe to both GPT 5.4 and Claude 4.6? Because 5.4 4.6 is a perfect 10/10. Meme: why do you think you could be smarter than Codex/Claude?

> went out for dinner > left claude loose on a "battle-tested" codebase > come back after 3h > 10 real bugs found the future of software engineer is feeling unreal

I was going crazy because I could not replicate TurboQuant. Turns out the community also had issues. The community quickly made adjustments to "make it work", but what they did not realize is that they reimplemented (most of) HIGGS in the process (full HIGGS would be even better)

My mental health for the day entirely depends on whether Claude Code woke up with a 200 IQ or a room-temperature IQ

got me for a sec 😅

so many familiar faces! @vinodg @bohanhou1998