Meta's Muse Glimmer Runs on One GPU, Claude Pushes a Riemann Bound to 67.2%, and OpenAI Ships a Restricted Cyber Model

Meta Returns to Open Weights with Muse Glimmer

The headline of the week is Meta's Muse Glimmer, a 30B dense, multimodal, agent-focused model released under a permissive Apache 2.0 licence, announced by Mark Zuckerberg and Alexandr Wang, with the larger Muse Spark 1.2 weights promised soon. The story that matters for ordinary users is deployment: quantised to around 18GB, Glimmer runs on a single consumer RTX 3090 using roughly 22-23GB of VRAM, with a 128K context and a bundled DFlash speculative-decoding drafter for snappy on-device responses. Early testers rate it strong for its size, if not the outright smartest, with rivals like Alibaba's Qwen expected to respond fast.

Claude Nudges the Riemann Hypothesis

Anthropic reported that an unreleased research Claude improved a longstanding lower bound on the fraction of zeta zeros on the critical line, from 41.6% to 67.2%. It did not solve the 160-year-old problem, but the workflow is the story: around 650 failed ideas, dozens of parallel subagents, thousands of numerical checks, and roughly 31M output tokens, all reviewed by human mathematicians. It's a vivid preview of AI as a tireless scientific collaborator.

OpenAI's Restricted Cyber Model

OpenAI launched GPT-5.6-Cyber under its Daybreak initiative, already used to find previously unknown bugs in open-source software and Chrome's V8 engine. Because a tool that finds vulnerabilities cuts both ways, access is limited to approved defenders with extra monitoring. Separately, Anthropic made Claude Sonnet 5's pricing permanent at $2/M input and $10/M output, another sign of a fierce price war.

Dyna-2 Learns Robotics from Human Video

Dyna Robotics introduced Dyna-2, a world-action model pretrained on 1 million hours of ordinary human video. The striking claim is transfer: learning from watching people helps it control unseen robots on new tasks, chipping away at robotics' biggest bottleneck, the cost of collecting real robot data.

Tiny Models, Big Ambitions

inclusionAI's Ling-3.0-tiny is an 8B mixture-of-experts model that activates only about 1.3B parameters per token, hitting 86-105 tok/s on consumer hardware. It underlines the week's clear trend: capable AI is escaping the data centre and moving onto laptops and phones.

Talking to Tools in Code

A widely shared paper argued that letting models call tools by writing code, rather than rigid JSON, matched or beat native JSON tool calling in 11 of 14 models, with the GPT-5.6 family gaining 10.6% on BFCL v4. Paired with speculative-decoding gains from systems like DSpark and DFlash, it's another quiet efficiency win for AI agents.

Hassabis and the 20-Year Cure

Google DeepMind's Demis Hassabis suggested all diseases could be cured within about 20 years. Enthusiasts pushed back sensibly: AI may accelerate discovery, but human clinical trials still set the pace.