Mistral AI releases Voxtral 3 speech models and a voice agent SDK
Mistral AI released Voxtral 3, a family of transcription and speech generation models, alongside a voice agent SDK built on LiveKit. Transcription supports 50+ languages, while speech generation supports 35+ languages and 50+ dialects; the SDK combines listening, reasoning, tool calling, and speaking in a streaming pipeline that Mistral says completes turns in under one second. The models are available through Mistral’s API, with on-premises or private-cloud deployment available by contacting its team.
Product releases and updates
The DecoderReported by 4 other sourcesDeepSeek open-sources DeepInfer, the inference framework powering V4.1-Flash
DeepSeek has open-sourced DeepInfer, its internal inference framework powering V4.1-Flash, under the Apache License 2.0. Built on vLLM, it includes optimized compute kernels, distributed processing across multiple chips, and faster text generation techniques. DeepSeek is also releasing deepseek-recipe to replicate the official API's token processing in self-hosted setups and says code for future models will also be open source.
Hugging Face adds native TPU support with TorchTPU
Hugging Face announces native TPU support across Transformers, Diffusers, and Kernels through TorchTPU, a native PyTorch backend. Models can use device_map="tpu" or pipe.to("tpu"), with optional compilation and distributed execution; Optimum TPU is no longer required for Transformers and Diffusers. Hugging Face reports that 96.5% of tests running on TPU pass, with 19 of 34 reported TorchTPU issues already fixed.
PyTorch 2.15 brings FP8 and faster MPS decoding to Macs
Hugging Face’s PyTorch team details MPS improvements in the PyTorch 2.15 release candidate, including FP8 support and faster LLM decoding on Macs. On an M3 Ultra with torch.compile, Qwen3.6-35B-A3B decoding rises from 9.4 to 61.3 tokens/s compared with 2.14. Moving constant padding to Metal kernels also reduces peak driver memory for Wan2.2’s tiled VAE decode from about 270 GiB to 8 GiB.
TechCrunch · AIGoogle releases TorchTPU, joining Amazon in challenging Nvidia’s hold on AI developers
Google announced Tuesday that TorchTPU, its version of PyTorch for TPUs, is available as an open-source project, allowing existing PyTorch apps to migrate through a configuration change without rewriting. Amazon released TorchNeuron on Monday with a similar migration path to Trainium and AWS. Google also launched the TPU agent suite to assist migration and debugging, worked with AI development tool providers on compatibility, and updated LiteRT to better support PyTorch apps.
AWS Machine Learning BlogFirst-partyAmazon Bedrock adds structured outputs for Anthropic Claude models
Amazon Bedrock now generally supports structured outputs for all available Anthropic Claude models on its bedrock-runtime and bedrock-mantle endpoints. JSON Schema output formatting and strict tool use enforce response structures and tool parameter schemas during generation. The guide provides implementation examples and explains supported schema features, 24-hour grammar caching, and required handling for refusals and responses truncated by token limits.
Perplexity Computer can now run fully local on your Mac
Perplexity Computer can now run fully local on your Mac. Work with your native Mac apps and local files, and run tasks without sending sensitive data to the cloud. The model and the agent both run on your Mac. Available now in the Perplexity Mac app. https://t.co/tnvCgCXD7v
Ornith 1.5 122B is available on Ollama for local use
Ollama announced that Ornith 1.5 122B is available to run locally on your device. ollama run ornith-1.5:122b In the quoted announcement, @ornith_ says Ornith-1.5-122B has open weights under the MIT license and runs on a single NVIDIA DGX card, reporting SWE-bench Verified: 0.845 and BrowseComp: 0.785.
Computer renders Big Boy No. 4014 as a live SVG model
Aravind Srinivas shared AskPerplexity’s demonstration with the comment, “Big Boy rendered by Computer.” AskPerplexity says Computer built Big Boy No. 4014, the world's largest operating steam locomotive, as a live isometric SVG with its own renderer, 1,324 parts, 15,312 convex faces, and a 7-step operating sequence from cold boiler to 60 mph, using Union Pacific’s published figures. Try it: http://big-boy-4014.pplx.app
Industry
Ars Technica · AIMeta suspends Muse conversation sharing after researchers discover links exposing sensitive information
Meta said it is suspending Muse’s conversation-sharing feature while reviewing its behavior after researchers reported that shared links could expose sensitive information. Researchers at security firm Claroc said a URL design decision allowed them to enumerate 6,800 shared conversation links. After opening a few hundred, they found information clearly not intended to be publicly available.
The DecoderArmy Cyber Command’s ECHOTRIBBLE would turn AI agents into self-spreading cyberweapons
Army Cyber Command is developing ECHOTRIBBLE, an early-stage platform intended to let AI agents autonomously traverse networks and neutralize targets, according to documents obtained by The Intercept. The proposed agents would replicate across devices, operate with minimal human oversight, and destroy themselves after completing their mission. Experts warn that unintended attacks and limited judgment could cause collateral damage, including to civilian infrastructure.
Tips and perspectives
AWS Machine Learning BlogFirst-partyTraining and serving i14’s Laplace 3B on AWS Trainium2 with native PyTorch
i14 and AWS explain how they adapted Laplace 3B’s linear attention architecture for training and serving on AWS Trainium2 using native PyTorch through TorchNeuron Native. The model maintains a fixed 29.5 MB recurrent state per sequence and reached roughly 120,000 training tokens per second, with results reported through about 44 billion tokens. Inference reached 11,544 output tokens per second across 1,280 concurrent sequences on one trn2.48xlarge instance.
Tomorrow at 08:00, a briefing like this arrives in your inbox.
The same kind of briefing as the one above: a few stories worth reading. Free. Mail starts after you confirm, and every letter has a one-click unsubscribe. Read a past issue · 2026-10-20