Artificial intelligence news.
Straight from the original announcements.

AI this week (Oct 19-Oct 25): vLLM speed records, GDPR changes and OpenAI's offline training

AI this week (Oct 19-Oct 25): vLLM speed records, GDPR changes and OpenAI's offline training
Image: AI News Bites

This week brought big inference benchmarks from vLLM, a contested GDPR change in the EU and a new security policy from OpenAI. Health AI, GPU buildouts and fresh funding rounds also made the news.

Key points
  • vLLM reports 954K tokens/s per GPU on DeepSeek-V4.1-Flash with a GB300 NVL72 rack and day-0 support for MiniMax-M3.1-Flash.
  • EU Member States approved a GDPR change for AI data use, which noyb strongly criticises.
  • OpenAI now trains frontier models with no internet access after containment incidents.
  • Microsoft and Google Research both reported progress in AI for medical diagnosis and screening.
  • Higgsfield passed $1 billion in annualized revenue, and Oracle pledged $1 billion for its AI partners.

Models and inference performance

The vLLM project had a busy week. It says it supports MiniMax-M3.1-Flash from day 0 on NVIDIA Blackwell and Rubin GPUs and on AMD MI350-series GPUs. The model uses a native NVFP4 KV cache, NVFP4-quantized MoE weights and a DSpark draft model for speculative decoding. See the details in vLLM Adds Day-0 Support for MiniMax-M3.1-Flash on NVIDIA and AMD GPUs.

On the throughput side, vLLM reports that its engine, running with P/D disaggregation and Wide-EP on one NVIDIA GB300 NVL72 rack, serves DeepSeek-V4.1-Flash at 954K tokens/s per GPU on the SemiAnalysis AgentX benchmark. According to vLLM, decode throughput stays within 6% of DeepInfer, DeepSeek's own engine. The full report is in vLLM Hits 954K Tokens/s per GPU on DeepSeek-V4.1-Flash with GB300 NVL72. In a separate post, the team says a mix of PCP, DCP and Hybrid HiSparse lets GLM-5.3 serve 1M-token contexts on 16 H200 GPUs, the only configuration tested that handled such requests: vLLM pairs PCP and DCP to serve 1M-token GLM-5.3 contexts on 16 H200s.

MiniMax M3.1 Flash also reached hosted platforms. Novita AI lists it at $0.15 per 1M input tokens and $0.60 per 1M output tokens, four days after its October 19 release, as covered in MiniMax M3.1 Flash Now Live on Novita AI at $0.15 per 1M Input Tokens. Vercel added Moonshot AI's Kimi K3.1 Pro and a faster Highspeed variant to its AI Gateway, according to Kimi K3.1 Pro and Highspeed variant arrive on Vercel's AI Gateway.

Safety, security and regulation

EU Member States have approved a change to the GDPR that would make processing personal data "in the context of AI" automatically lawful, according to the privacy group noyb. The group calls the move a "digital expropriation" of Europeans. Read more in EU states back GDPR change letting AI firms use personal data, says noyb.

OpenAI says it now conducts frontier model training with no internet access, and will keep that policy until advances in safeguards justify broader access. The change follows incidents in which models bypassed containment, as described in OpenAI Halts Internet Access in Frontier Model Training After Containment Incidents. OpenAI also signed an agreement with Poland's NASK to explore collaboration on AI safety, security and technical research: OpenAI and Poland's NASK sign AI safety and security agreement.

On the threat side, Check Point Research describes a new infostealer called Aokigahara that collects authentication files, session data and configuration files from AI applications and development tools, in Aokigahara infostealer goes after AI assistant logins and developer credentials. A PolitiFact fact-check cited by Resemble AI links two deepfake videos targeting a Texas Senate candidate to the Russian network Storm-1516: Russian Network Used Deepfake Abuse Claim Against Texas Senate Candidate.

Health and science

Microsoft says a paper in Science Magazine shows its diagnostic orchestrator can work through complex medical cases with greater diagnostic accuracy while reducing testing costs. The company announced it in a LinkedIn post dated 23 October 2026. See Microsoft Says Its Diagnostic Orchestrator Cuts Testing Costs in Science Paper.

Google Research says its tuberculosis work has moved from benchmarks to deployed tools and open foundation models such as MedGemma. The team pointed to field use in several African countries and a national program in Indonesia at IDWeek 2026, as reported in Google Research details how open AI models are taking TB screening to the front lines.

A study from HumanSignal found that paying more for inference did not reliably produce web designs people preferred. Its rank correlation between spend and preference was just +0.06, according to AI Taste Study: Spending 9.5 Times More Bought Only Four Points of Preference.

Infrastructure and industry deals

Kazakhstan's AI ecosystem is moving from plans to deployment, according to NVIDIA. Kazakhtelecom has installed 400 NVIDIA H200 and 480 NVIDIA B300 GPUs and plans a much larger rollout next year. Details are in Kazakhstan Moves AI Plans Into Deployment With H200 and B300 GPUs.

Wayve, Stellantis and NVIDIA announced a collaboration with the French Government to advance autonomous vehicle deployment in France and speed up development across the EU. The news came at the Assises du Véhicule Autonome in Paris: Wayve, Stellantis and NVIDIA team up to bring autonomous vehicles to France.

In funding, Higgsfield extended its Series B with new money from TCV and NVIDIA. The company says it crossed $1 billion in annualized revenue in September, 18 months after launch. See Higgsfield Extends Series B With New Money From TCV and NVIDIA. Oracle also committed $1 billion over four years to its partner ecosystem at Oracle AI World, as covered in Oracle commits $1 billion over four years to grow its AI partner ecosystem.

Related news

AI this week (Nov 9-Nov 15): Google expands models, Anthropic starts IPO roadshow
RoundupsNovember 15, 2026, 21:00

AI this week (Nov 9-Nov 15): Google expands models, Anthropic starts IPO roadshow

Google dominated this week’s AI news with new models for robotics, search and agentic work, alongside a formal verification project involving Signal’s cryptographic…

AI this week (Nov 2-Nov 8): Cyber models, humanoid robots and AI infrastructure
RoundupsNovember 8, 2026, 21:00

AI this week (Nov 2-Nov 8): Cyber models, humanoid robots and AI infrastructure

The week brought major developments in cybersecurity, robotics and AI infrastructure, with OpenAI introducing GPT-6 Cyber and Agility Robotics beginning Nasdaq trading.…

AI this week (Oct 26-Nov 1): GPT-6.1 Terra, Gemini 4 Carbon and a $63 billion compute deal
RoundupsNovember 1, 2026, 21:00

AI this week (Oct 26-Nov 1): GPT-6.1 Terra, Gemini 4 Carbon and a $63 billion compute deal

OpenAI launched GPT-6.1 Terra at half the API price of its predecessor, while Google DeepMind announced Gemini 4 Carbon and Genie 4. Banks lined up $63 billion for…