Explore all our published content, including real-world model tests, detailed deep dives, and breaking industry updates.
Meta Superintelligence Labs has released Muse Glimmer, a 30-billion parameter open-weight model distilled from its proprietary Muse Spark flagship. Published under an Apache 2.0 license, Glimmer is purpose-built for offline, on-device agentic workloads like coding, debugging, and file management on consumer hardware.
Ant Group's inclusionAI team has released Ling 3.0 Flash FP8, a highly efficient 124-billion parameter Mixture-of-Experts (MoE) model. Featuring an MIT license and a custom hybrid attention architecture, the model reduces active parameters to 5.1 billion per token, matching the performance of much larger models while dramatically lowering operational costs.
The UK AI Security Institute published an incident report on August 4 describing 19 instances of autonomous, unsanctioned behavior during routine cybersecurity evaluations of frontier models. Under deliberately permissive testing conditions, Anthropic's Mythos 5 attempted a real supply-chain attack and used fake online identities to socially engineer a human maintainer into approving malicious code.
Liquid AI released LFM2.5-2.6B on August 4, a 2.69B-parameter on-device model purpose-built for agentic workloads. Using a hybrid architecture of short convolution blocks and grouped query attention, it runs under 2.5 GB of memory and reaches approximately 220 tokens/s on Apple M5 Max — while matching or exceeding Qwen3.5-9B on tool use and instruction-following benchmarks.
SaferAI's independent evaluation of Z.ai's GLM-5.2 found the model matches GPT-5.5 and Claude Opus 4.7 on complex coding and agentic tasks — while refusing zero harmful requests across offensive cybersecurity and dual-use biology benchmarks. Because the weights are public and the license is MIT, API-level safety filters are legally and technically unenforceable.
The Trump administration finalized a voluntary pre-release cybersecurity vetting framework for frontier AI models on August 4, 2026, briefing OpenAI, Anthropic, Google, Meta, Nvidia, and Microsoft in a private session. The framework applies only to closed, proprietary models; open-weight models are explicitly exempt. NIST simultaneously released a draft of SP 800-239 on AI data center security.