DeepSeek V4 Pro 0813 Goes GA — But Independent Benchmarks Are Missing
DeepSeek quietly promoted its flagship model, V4 Pro 0813, to general availability on August 12 — no blog post, no press release, just a version bump on the API endpoint. The model is a mixture-of-experts system with 1.6 trillion total parameters and 49 billion active per token, making it one of the largest publicly available language models in existence.
DeepSeek claims benchmark improvements of up to 49.9 percentage points over prior versions, but as TechTimes notes, no third-party evaluator has replicated the scores yet. The model exposes three operating modes: standard non-thinking, high-effort reasoning, and a maximum-effort V4-Pro-Max configuration.
Two architectural innovations stand out: Compressed Sparse Attention and Heavily Compressed Attention, which DeepSeek says reduce single-token inference compute to 27% and KV cache to 10% of what V3.2 required at the 1M-token context setting. The model supports a full 1,048,576-token context window with up to 384,000 tokens of output.
There's a catch on pricing. Current rates — $0.44 per million uncached input tokens and $0.87 per million output tokens — remain competitive, but on August 17, DeepSeek will introduce peak and off-peak pricing that could raise rates by up to 12x. Notably, V4 Pro still lacks vision support, a trade-off the community remains divided on.
Gemini Crosses 1 Billion Monthly Users
Google CEO Sundar Pichai announced on August 11 that the Gemini app has surpassed 1 billion monthly active users, making it Google's fastest-growing product ever and its 14th service to hit the billion-user milestone — joining the ranks of Search, Gmail, Android, and YouTube.
The growth trajectory has been striking: Gemini went from 400 million users in May 2025 to 900 million by May 2026, then added the final 100 million in under a month. Google credits deep integration across Android, Chrome, Google Search, and Workspace for the acceleration.
Usage patterns reveal that 63% of Gemini users interact through voice, while the platform generates more than 150 million images daily. ChatGPT reportedly reached 1 billion users in June, but some analysts note Google's metric tracks monthly actives versus weekly, making direct comparisons tricky. The AI assistant race is officially a two-horse contest at the billion-user scale.
Meta Drops Muse Glimmer: A 30B Open-Source Agent That Runs on Consumer Hardware
Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter multimodal model under an Apache 2.0 license, designed specifically for local agentic workflows. The model was distilled from the larger Muse Spark using logit distillation, then further refined with agent-focused training, supervised fine-tuning, and reinforcement learning.
The headline feature is efficiency: 4-bit quantization compresses Muse Glimmer under 20 GB, fitting it in 24 GB of VRAM with only 1% quality degradation. Combined with DFlash 16-token block speculation, Meta claims a 3.1x decode speedup on RTX 5090 hardware. It supports a 131K context window and over 100 languages.
The model is available now through Hugging Face with day-one support for Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM, and SGLang. For developers building local coding agents, LLM-as-a-judge evaluation pipelines, or privacy-sensitive agentic tools, Muse Glimmer is positioned as a compelling alternative to cloud-only frontier models.
Manus Returns to Independence After Beijing Blocks Meta's $2B Acquisition
AI agent startup Manus announced on August 11 that it will resume operating as an independent company, following Beijing's order to unwind Meta's $2 billion acquisition. The deal, announced in December 2025 and closed on December 29 of that year, was struck down by China's National Development and Reform Commission in April, citing foreign investment rules.
The forced divestiture reflects China's crackdown on “Singapore-washing” — a practice where Chinese AI companies relocate to Singapore to access looser regulations and foreign funding. Manus, originally founded in China in 2022, had moved its base to Singapore before the Meta deal.
The separation carries practical consequences: user data created from December 29, 2025 onward is slated for deletion in certain jurisdictions. As Bloomberg reports, Manus must now chart its own course in an increasingly competitive AI agent market — without Meta's resources but also without the geopolitical baggage.
EU AI Act Enforcement Kicks In With New Transparency Rules
As of August 2, the European Commission's AI Office, together with national authorities, began enforcing the AI Act with new transparency requirements. Chatbots and interactive AI systems must now inform users when they're interacting with AI rather than a human, and deepfakes must carry clear labels.
The enforcement timeline remains staggered: high-risk AI system provisions have been postponed to December 2027 for standalone systems and August 2028 for those embedded in products. However, the law already bans AI systems that generate non-consensual sexual or intimate content and child sexual abuse material.
Meanwhile, the global regulatory landscape remains fragmented. The US still lacks a single federal AI law, operating instead through a patchwork of state bills, agency actions, and sector-specific oversight. For companies operating across borders, compliance is becoming an increasingly expensive multi-jurisdiction puzzle.