A 12-million-token memory, a tiny model that punches up, and an AI that discovers new maths

SubQ: a 12-million-token context window

Miami startup Subquadratic launched SubQ, the first commercial large language model built on a fully sub-quadratic architecture, offering a native 12-million-token context window. Its Subquadratic Selective Attention scales roughly linearly rather than quadratically, claiming 52x the speed of FlashAttention at 1M tokens and up to 1,000x less compute at full length. It tops long-context retrieval benchmarks but trails leaders on broader coding tasks, sparking heated debate.

ZAYA1-8B: frontier intelligence density on AMD

Zyphra released ZAYA1-8B, an Apache 2.0 mixture-of-experts reasoning model with 8.4B total but only ~760M active parameters. It matches or beats far larger open models on maths, coding and reasoning, and was trained entirely on a cluster of 1,024 AMD Instinct MI300X accelerators, proving the non-dominant hardware stack works at scale.

AlphaResearch: an AI that discovers algorithms

AlphaResearch is an autonomous research agent that proposes ideas, verifies them in code, and refines direction using both execution-based rewards and a simulated peer-review environment. It set a best-known result on the circle-packing problem, beating humans and prior systems like AlphaEvolve, and outperformed rival agents on six of eight open problems.

NVIDIA opens its physical-AI stack

NVIDIA released five open model families (Nemotron, Cosmos, Alpamayo, Isaac GR00T, Clara) alongside massive open datasets: 10 trillion language tokens, 500,000 robotics trajectories, 455,000 protein structures and 100TB of vehicle sensor data, plus open training frameworks.

AGIBOT World 2026 open robot dataset

AGIBOT open-sourced a richly annotated, multi-sensory real-world robot dataset (RGB-D, tactile, lidar, IMU, joint states) collected free-form in homes and commercial spaces to improve generalisation for embodied AI.

Neuro-symbolic AI cuts energy 100x

Tufts researchers combined neural networks with symbolic reasoning, hitting 95% success on the Tower of Hanoi versus 34%, cutting training from 36+ hours to 34 minutes and using around 1% of the energy.

AI for rare-disease diagnosis

A Nature-reported AI system fuses clinical data, genetics and the medical literature to suggest rare-disease diagnoses with transparent reasoning a clinician can verify, potentially shortening the years-long diagnostic odyssey.

Gemma 4 open reasoning models

Google's Gemma 4 family ships open weights under Apache 2.0, built for multi-step reasoning and agentic workflows with native tool-calling and structured output, continuing the intelligence-per-parameter theme.