AI News Digest: Meta's Muse Glimmer, Anthropic Math Bounds, and Agent Tooling

Meta released Muse Glimmer, an open-weight 30B dense multimodal model optimized for local always-on agents and long-horizon tasks. Built with hybrid attention, a dedicated perception encoder, and a bundled DFlash speculative decoding drafter, the model can run on consumer hardware like a single 24GB RTX 3090 when using 4-bit quantization and efficient sliding-window key-value caches. Day-one ecosystem support includes vLLM, llama.cpp, Ollama, and Unsloth.

In other frontier developments, an unreleased research version of Anthropic's Claude improved a lower bound related to the Riemann Hypothesis by increasing the fraction of zeta zeros on the critical line from 41.6% to 67.2% using extensive proof iteration. Meanwhile, OpenAI introduced GPT-5.6-Cyber under restricted access for authorized defensive vulnerability research, and Anthropic made its introductory pricing for Claude Sonnet 5 permanent.

The broader AI community focused heavily on agent harnesses, inference optimization, and tool interfaces. Evaluations showed that treating tools as executable Python code rather than JSON schemas improves performance, while token efficiency experiments significantly reduced the cost of browser automation. Additionally, community discussions highlighted various local hardware setups and quantization strategies for running models like DeepSeek V4 Flash and Qwen derivatives.