Alibaba Drops Qwen 3.8-Max: The Largest Open-Weight Model Yet
Alibaba's Qwen team released Qwen 3.8-Max on Sunday — a 2.4 trillion parameter Mixture of Experts model with 95 billion active parameters and a 1 million token context window. It's the most capable model in the Qwen family to date, accepting text, image, and video input.
What makes this release stand out isn't just the scale — it's the commitment to openness. Alibaba plans to release the full open weights on Hugging Face and ModelScope next week, marking the first time the company has open-sourced a model at this tier. The API is already live on QwenCloud at $2/$6 per million tokens (input/output).
The coding capabilities are particularly impressive. In one demonstration, the model ran autonomously for 16 days on an open-source CLI project, producing 265 commits and 127 pull requests — iterating through its own code, tests, previews, and logs without human intervention. Qwen 3.8-Max also ships with a smaller 27B sibling optimized for coding and collaborative workflows.
Source: MarkTechPost, Quartz
EU AI Act Enforcement Goes Live — Fines of Up to 3% of Global Revenue
As of August 2, the European Commission can officially fine companies that build general-purpose AI models. The one-year grace period on the EU AI Act's GPAI provisions has ended, and the Commission's AI Office now has the power to request documentation, run technical evaluations, demand compliance measures, restrict models from the EU market, and issue fines of up to 3% of global annual turnover or €15 million, whichever is higher.
The scope is deliberately extraterritorial: every company that makes a general-purpose AI model available in the EU is covered, regardless of where it's headquartered. That puts OpenAI, Anthropic, Google, Meta, and every other major AI lab squarely in the crosshairs. Companies must now demonstrate how their models were trained, what data was used, and what risk-mitigation measures are in place.
This is a significant milestone — the EU has moved from publishing rules to actively enforcing them, and the rest of the world is watching to see how aggressively the AI Office wields its new powers.
Source: SpaceDaily, Quartz
DeepSeek V4 Flash Exits Preview, Outperforms Its Own Pro Model
DeepSeek officially released V4 Flash 0731 last week, and the numbers are turning heads. The budget model scored 82.7% on Terminal Bench 2.1, a widely-used agent benchmark — beating DeepSeek's own V4 Pro Preview, which managed only 72.1%. That's a 14.7% advantage for a model that costs a fraction of the price.
The pricing is aggressive: $0.14 per million input tokens and $0.28 per million output tokens, making it one of the cheapest capable models on the market. For context, the median cost for comparable open-weight models is $0.58/$2.20. V4 Flash also generates output at 122.7 tokens per second with a time-to-first-token of just 1.32 seconds.
This release continues DeepSeek's pattern of commoditizing AI inference. When your budget model beats your premium model on the benchmarks that matter most for real-world agent use cases, it raises uncomfortable questions about the value proposition of expensive frontier models.
Source: Artificial Analysis, Flowtivity
Genspark Open-Sources GenOffice: A Full AI Office Suite
Genspark released GenOffice under the Apache 2.0 license — a complete AI-native office suite for macOS and Windows that includes a word processor, spreadsheet editor, presentation tool, and PDF editor. It's free, ad-free, and runs as a desktop application rather than in the cloud.
The backstory is remarkable: one engineer built the alpha version in a single week using $10,000 worth of AI tokens. GenOffice reads and writes standard file formats (.docx, .xlsx, .pptx, .pdf) and has Genspark's Super Agent built into every editor, capable of researching topics, analyzing data, writing sections, and building entire slide decks.
While still in alpha, GenOffice is a provocative proof of concept for the "AI-built software" thesis. If one person and $10K in compute can produce a functional office suite in a week, the implications for traditional software development are hard to ignore.
Source: Genspark Blog, GitHub
The AI Talent Wars Are Getting Personal
The battle for AI researchers is intensifying, and Anthropic is winning. According to a recent Axios report, engineers at OpenAI are eight times more likely to leave for Anthropic than the other way around, while at Google DeepMind the ratio is nearly 11:1 in Anthropic's favor. The company boasts an 80% retention rate for employees hired over the past two years.
The most notable recent moves include Noam Shazeer — co-author of the landmark "Attention Is All You Need" paper and Gemini co-lead — who left Google for OpenAI, and John Jumper, the Nobel Prize-winning AlphaFold researcher, who departed for Anthropic.
Google has been hit hardest, with multiple employees citing the company's Pentagon deal as a reason for leaving. Meta, meanwhile, has invested heavily in recruiting for its superintelligence division, only to watch several high-profile hires quickly leave for competitors — some to OpenAI.
Source: Axios