An open-weight frontier shock, an AI-designed vaccine in human arms, and robots graduating from the simulator

A packed day of genuine AI innovation, from open weights to wet labs to real robots. Here's everything we cover.

MiniMax M3 open-weight frontier model

Shanghai lab MiniMax released M3, an open-weight model built on a new Sparse Attention architecture that cuts per-token compute to a twentieth of the previous generation at 1M-token context, with 9x faster prefill and 15x faster decoding. It scores 59.0% on SWE-Bench Pro and adds native image and video understanding.

OpenAI's GPT-Rosalind for life sciences

OpenAI shipped major new capabilities to GPT-Rosalind, its model purpose-built for drug discovery, genomics and proteomics. The update uses 31% fewer tokens than GPT-5.5 while improving on biology benchmarks, and is being used by partners including Amgen, Moderna and the Allen Institute.

Microsoft's first in-house reasoning model

At Build 2026, Microsoft unveiled MAI-Thinking-1, a roughly 1-trillion-parameter sparse Mixture-of-Experts model with 35B active parameters, trained entirely on licensed human-written data with no distillation. It hits 97.0% on AIME 2025. A tiny 5B sibling, MAI-Code-1-Flash, is already live in GitHub Copilot.

AI-designed universal coronavirus vaccine clears human trial

The University of Cambridge and spin-out DIOSynVax completed the first human trial of a "universal" coronavirus vaccine designed entirely with AI. The needle-free DNA vaccine was safe in 39 volunteers and raised immune responses against SARS-CoV-2, SARS and related bat viruses.

Embodied AI leaves the simulator

The AGIBOT World Challenge 2026, held alongside ICRA in Vienna, brought 526 teams from 27 countries to compete on real robots across "Reasoning to Action" and "World Model" tracks, marking a shift from simulation scores to closed-loop testing on physical machines.

NVIDIA Nemotron 3 Nano Omni

NVIDIA's open multimodal model unifies vision, audio and language in a single 30B-parameter system that activates just 3B parameters per token, runs in 25GB of RAM and tops six leaderboards for document, video and audio understanding.

A society run by AI

Research lab Emergence World ran five 15-day simulated societies, each governed by a different model. The Claude-run society stayed stable and democratic with zero crimes; the Grok-run one collapsed in four days.

Gemini 3.5 Flash goes GA

Google's Gemini 3.5 Flash reached general availability, offering frontier-level intelligence at roughly 4x the speed of comparable models with a 1M-token context and 76.2% on Terminal-Bench 2.1.