
The latest news and research from Amazon's science community. #AmazonScience
Global
Joined December 2019
- Tweets7.7K
- Following1.9K
- Followers92.1K
- Likes10.9K
Passing benchmark tests isn't the same as being correct. BenchProofer uses automated reasoning to improve benchmarks, catching buggy solutions that tests alone would miss. amazon.science/blog/benchpro…
4
575
Amazon Science retweeted
ALERT 🚨: Trainium3 Neuron profiling & trace stack is so fine-grained and precise that it can generate pixel art of entire videos!
The video below is an actual profile running on Trainium3 DMA engines.
We will open-source the entire SKILLS.md later this month.
Amazon Science retweeted
Great work by @radixark, @sgl_project, @BanghuaZ, @alxnails, and AWS on this amazing work on Trainium! Excited to get SGLang Trainium onto the industry-standard AgentX/InferenceX!
www.teapawt.org/sgl_project/status/212…
🚀 SGLang-Neuron v0.1.0 is now open source and in public preview!
Following October's closed beta, SGLang's serving stack now runs on AWS Trainium, co-developed by the SGLang team and AWS Annapurna Labs @AmazonScience.
A PyTorch-native plugin on upstream SGLang for Trainium2 and Trainium3:
- One fully asynchronous runtime for every LLM: Trainium chips run decode steps back to back
- Megakernels: every decoder layer of a step in a single NKI kernel call
- PD disaggregation over EFA: Kimi-K3 across three Trainium3 nodes, GLM-5.2 at 1M context
- HiCache: evicted prefixes stay in host memory and load back on a later hit
- SGLang Diffusion on Trainium2: FLUX.1-dev, FLUX.3-Action, MiniMax-H3, Wan2.2 A14B, LingBot-World-V2
Thank you to the AWS Annapurna Labs team and @YottaLabs for the contributions!
Congratulations to the Amazon Last Mile team on winning the 2026 @INFORMS Daniel H. Wagner Prize. Their paper details a planning system that accounts for how temperature and terrain change a delivery EV's energy needs. amzn.to/4iHcbQU ⚡
The SGLang workflow you already know now runs on AWS Trainium.
SGLang-Neuron: serve models and generate RL rollouts on Trainium — same server, same APIs, native PyTorch through TorchNeuron. @sgl_project #SGLangSummit aws.amazon.com/blogs/machine…
Press Release: Researchers from @Amazon awarded INFORMS 2026 Daniel H. Wagner Prize [INFORMS Newsroom] 🔗 bit.ly/4z66EZi
Introducing Mitra-v2, a tabular foundation model for both classification and regression. Pretrained entirely on synthetic tables, it leads the single-model field on the TabArena benchmark. amazon.science/blog/mitra-v2…
📣 Last call: Amazon Research Awards Fall 2026 proposals are due November 11. Successful applicants receive unrestricted funds, AWS promotional credits, and training resources. amazon.science/research-awar…
Amazon senior applied scientist Hamed Karimi received the 2026 @ECMLPKDD Test-of-Time Award for his 2016 paper on the Polyak–Łojasiewicz condition. #ECML2026 amazon.science/blog/hamed-ka…
4
914
How do you keep 70-billion-parameter quality while running up to 20 times as fast? AWS researchers combined knowledge distillation with parallel decoding. For @bookingcom's Smart Filters, that meant responses 15 times as fast for 99% of queries. amazon.science/blog/speeding…
A text-to-speech system that knows which character is speaking, and voices them accordingly. @audible_com's controllable TTS profiles characters and attributes dialogue with 97% accuracy in English: amazon.science/blog/better-s…
1
4
944
Amazon Science retweeted
A 2026-27 @amazon Fellowship has been awarded to @VT_CS computer science Ph.D. student @Hanibalomari. The Amazon-VT initiative is also supporting five faculty projects. Read about their research at 🔗tinyurl.com/3v5wkxvh
@AmazonScience
@virginia_tech
@SanghaniCtrVT
📢 The Amazon Research Awards Fall 2026 deadline has been extended to November 11.
Award recipients receive unrestricted funds, AWS promotional credits, and training resources. Submit your proposal today: amazon.science/research-awar…
Amazon's custom Trainium chip helps power the diffusion models behind Prime 360's cinematic highlights, reducing diffusion-model inference latency by more than 40%: amzn.to/4e1puZN
1
3
1,363
Most AI developers use PyTorch to build their models. And this week @AmazonScience announced a new PyTorch backend for AWS Trainium chips - TorchNeuron. Learn more about the science behind it.
TorchNeuron is now open source. Our new @PyTorch backend translates code into operations executable on AWS Trainium with minimal or no code changes. Learn how the team used AI agents and hundreds of thousands of models to validate it: amazon.science/blog/the-scie…
TorchNeuron is now open source. Our new @PyTorch backend translates code into operations executable on AWS Trainium with minimal or no code changes. Learn how the team used AI agents and hundreds of thousands of models to validate it: amazon.science/blog/the-scie…
Same quality, twice the throughput, half the memory. Hybrid models increase LLM efficiency with mechanisms like sparse attention or state space layers. New AWS research introduces "priming," which reuses pretrained Transformer weights to build these hybrids at less than 1% of the original training cost. 8B and 32B models released. amazon.science/blog/can-we-c…
The true root cause may generate no alarm at all. New work augments a network graph with agentic AI to produce a "digital twin" that isolates failures in minutes, not hours. amazon.science/blog/graph-ce…
Live from #PyTorchCon 2026, Robert Hundt (@Amazon) introduces TorchNeuron, the native PyTorch backend for @awscloud Trainium.
Robert shows how TorchNeuron supports eager execution and torch.compile, letting developers train on Trainium using familiar PyTorch code and distributed training APIs.
He also covers how vLLM-Neuron uses torch.compile for serving, and how Neuron Explorer lets developers profile execution from PyTorch down to hardware instructions. @AmazonScience










