Meta bets on open models for single graphics cards, OpenAI opens up to advanced cybersecurity

Meta’s chief has issued a manifesto-letter announcing the release of two new open-weight artificial intelligence models: Muse Glimmer and Muse Spark 1.2. Muse Glimmer is a dense multimodal model with 30 billion parameters, distributed under the Apache 2.0 licence and designed for local agentic use, programming, and the role of automated evaluator of other models, with a context window of 131 thousand tokens and support for more than 100 languages. The most notable feature concerns efficiency: through 4-bit quantisation, the model is compressed to under 20 gigabytes, allowing it to run on a single consumer-grade graphics card with a 3.1-fold speed increase. The news is confirmed by two distinct sources: Meta’s corporate statement as reported by AI Weekly and the chief’s manifesto-letter as reported by ANSA.

On the opposite front of the artificial intelligence industry, OpenAI expanded its Daybreak initiative on 10 August 2026, introducing two distinct access tiers. The first, Daybreak Blue, is based on the GPT-5.6 Sol model with cybersecurity safeguards removed at system level. The second, Daybreak Red, grants selected parties access to a dedicated model, GPT-5.6-Cyber, capable of answering advanced security queries normally blocked by standard systems. This information currently comes from a single source, AI Weekly, which attributes it to a corporate statement by OpenAI; no further independent confirmations appear in the available excerpts.

The two strategies outline opposite approaches to the same phase of the industry: Meta is betting on the wide distribution of open models that can run on limited hardware, while OpenAI is building a graduated, selective access system for capabilities that are normally restricted, explicitly tied to the theme of cybersecurity. The available excerpts do not specify the criteria by which OpenAI selects the parties admitted to the Daybreak Blue and Red tiers, nor whether there is a public application process to access them.

What remains, as verifiable data, are the technical figures declared by Meta — 30 billion parameters, compression to under 20 gigabytes, tripled speed on a single GPU — and the distinction between the two access tiers announced by OpenAI on the same day.