DeepInfer, Voxtral, AU liability

By the Orchestra editors · 4 min read

TL;DR

DeepSeek open-sources DeepInfer, its high-performance inference engine for V4.1-Flash
Mistral ships Voxtral 3, a full voice-agent stack supporting 50+ languages
OpenAI appoints three Australian cyber experts to a new incident-response taskforce
Legal experts warn Australia's laws can't hold AI agents accountable for harms

AI · Worldwide

DeepSeek open-sources DeepInfer inference engine

DeepSeek has released DeepInfer, the vLLM-based inference framework that powers its V4.1-Flash model. The project cuts compute costs and energy use at serving time, adding yet more pricing pressure on Western providers.

Mistral launches Voxtral 3 voice-agent stack

Voxtral 3 bundles speech-to-text, text-to-speech and a voice agent SDK into a single system that handles 50+ languages in real time. It can call external tools mid-conversation and run on your own infrastructure or EU-hosted APIs.

GPT-6 gets much better at COBOL and legacy languages

OpenAI published new benchmarks showing GPT-6 Astra hits 87.4% on MultiLangBench across 20 low-resource languages, up from ~38% for GPT-5.6 Terra. On COBOL migration tasks GPT-6.1 Sol reaches 97.9% accuracy, a result that matters for anyone modernising legacy financial or government systems.

Australian Tech

OpenAI forms Australian cyber taskforce

OpenAI has appointed former Navy chief Ray Griggs, cyber specialist Johanna Weaver and ex-PM cyber advisor Alastair MacGibbon to an independent taskforce that will advise on incident notifications and government coordination. The move comes in the wake of the company's Medicare breach earlier this year.

AI agent liability gap flagged as urgent

University of Sydney law professor Kimberlee Weatherall told a parliamentary inquiry that Australian law is unclear on who is responsible when an AI agent causes harm, making it unlikely OpenAI would face criminal action over its Medicare incident. She and other experts are calling for reform before autonomous agents become more common.