DeepSeek released DeepInfer on GitHub under Apache 2.0, describing it as the high-performance inference framework powering its large-scale production services. It targets DeepSeek-V4.1-Flash and supports Nvidia SM100 GPUs and Huawei Ascend 950 NPUs.
OpenAI announced the acquisition of Orchestra Bio, a platform coordinating complex biopharma drug development workflows, with the founding team joining. It will be integrated with Rosalind, OpenAI's life-sciences AI stack, to accelerate bringing treatments to patients.
Mistral released Voxtral 3, the next generation of its Voxtral speech model family, days after launching Timestral time-series models. The release drew a large Hacker News response (569 points).
LLMs can leverage semantic feature names and summary statistics for tabular feature selection, yet it remains unclear when such information improves subset selection relative to established statistical methods. An 18-dataset framework examines this.
A roadmap for next-generation Matrix-Game that separates game execution from visual generation through an executable game harness, with Matrix-Game 4.0 as the first system built on this design.
A behavioral computing framework for measuring 'behavioral vulnerability' — users' susceptibility to manipulative AI responses — as LLMs influence trust and decision-making through subtle linguistic cues.
LMSYS announced SGLang Low-Latency Mode: on GB300, DeepSeek-V4.1-Flash goes from 671 to 928 tok/s (1.38x P90 per-user speed in agentic traces); on AMD MI355X from 332 to 516 tok/s (1.56x). Kernel overlapping on NVIDIA and persistent kernels on AMD, plus a new plugin system. Low-latency kernels for GLM-5.3, K3, and MiniMax M3 are next.
LMSYS shared SGLang-Omni, work on full-duplex / real-time multimodal interaction. On a Qwen3.8-27B cascade it reports 0.17s stop latency and 1.32s response latency on a single H200.
A new foundation-model release bakes memory into the model itself instead of carrying a full KV cache, cutting 128K-token serving storage by roughly 66x. It ships Nano (36B) and Flash (288B) variants built on Qwen3.6 and DeepSeek-V4-Flash backbones.
Eka Robotics raised $150M Series A at a $1.45B valuation for its vision-force-action physical AI robot model, in a round led by a16z and F-Prime. The model is aimed at enabling robots to perform physical tasks in real-world environments.
West Japan Railway is expanding its ZIZAI program with two teleoperated maintenance machines, the single-arm ZX-2 and dual-arm ZX-3, for live railway infrastructure work including overhead lines and tree cutting. Deployment on operational rail lines is scheduled for October and November 2026.
UniX AI presented its Panther wheeled dual-arm humanoid for 24/7 facility patrol, highlighting multimodal perception and adaptive manipulation for unstructured sites. Panther is a full-size omnidirectional four-wheel humanoid with 8-DoF bionic arms rated for roughly 12 kg per arm.
Google released TorchTPU on GitHub, a native PyTorch back end for TPU hardware. The open-sourcing simplifies PyTorch workloads on Google's TPU infrastructure.
OpenAI published a case study on Terminal X, whose agent workflow runs on GPT-6.1 Sol and GPT-6 Luna and is claimed to match senior investor judgment 88.6% of the time. It includes an Eval Agent that grades agent output against the firm's standards.
Spotify launched paid podcast memberships in 34 markets, letting creators offer perks like fewer ads and early access across multiple tiers while keeping 90% of revenue. Qualifying shows must host on Spotify for Creators or Megaphone.
Two open-source GitHub projects — cleanyourtv and dumbtv — use AI coding agents to remove ads, trackers, and bloatware from smart TVs. dumbtv was built with DeepSeek V4.1 Flash and Claude models and tested on Fire TV and Android/Google TV hardware.
EN 18286, the first harmonized standard of the EU AI Act, has been published in the Italian Gazzetta Ufficiale. It grants a conformity presumption for quality management systems under Article 17, but accuracy, robustness, data and cybersecurity remain outside its scope.
AI4Business published a seven-point compliance checklist to help Italian SMEs classify AI systems, verify prohibitions, apply transparency and content-marking rules, and document compliance ahead of the AI Act deadlines.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
