August is traditionally the month when Europe slows down. The LLM world, evidently, does not: between August 5 and 14, 2026, Muse Spark 1.2 from Meta, Grok 4.6 from xAI, DeepSeek V4 Pro 0813, the weights of Qwen 3.8 from Alibaba, and GLM-5.3 from Zhipu all arrived one after another. Five frontier releases in ten days.
But the interesting data point is not the quantity. It is the geography of the licenses.
Two worlds, two strategies
On one side, the US labs, increasingly closed. Grok 4.6 arrived on August 12: closed weights, API at $2/$6 per million tokens, no download. And Meta - yes, Meta, the company behind Llama, for years the flag-bearer of US open source - has now completed its conversion: Muse Spark 1.2, released on August 5 alongside the coding agent Muse Code, is a proprietary model. No weights, no self-hosting, no fine-tuning. The Llama era is officially over. (As a consolation prize, Muse Glimmer arrived on August 10, a 30B open model under Apache 2.0 designed to run locally - appreciated, but not the flagship.)
On the other side, the Chinese labs, increasingly open - and increasingly competitive:
- DeepSeek V4 Pro 0813 (August 12-13): the 1.6 trillion parameter flagship (49B active per token, Mixture-of-Experts architecture) went into general availability with weights on Hugging Face under MIT license, the most permissive one available. A 1 million token context window, and prices that - despite a recent upward revision introduced on August 17 with peak and off-peak tiers - remain a fraction of US frontier model pricing.
- Qwen 3.8 from Alibaba: for the first time in the history of the Qwen family, even the Max-tier model (2.4 trillion parameters) was made downloadable, followed on August 14 by its smaller sibling at 27B under Apache 2.0. Until recently the industry rule was “download the small model, the best one stays behind the API”: Alibaba broke it.
- GLM-5.3 from Zhipu (August 14): same base as GLM-5.2, but with post-training aggressive enough to claim a +50% in coding and the title of best open-weight model for programming. Weights were promised pending security reviews.
- And at the end of July there was already Kimi K3 from Moonshot: 2.8 trillion parameters, open, at a fraction of frontier model pricing.
The pattern is clear: those who lead the revenue rankings sell access, those who chase give away the weights. If you can’t beat them on the product… go open source.
We have seen this before: the OS wars
Anyone who has worked in IT long enough has a sense of déjà-vu. In the 1990s and 2000s the script was identical, just with different actors.
On one side the proprietary operating systems - Windows, the commercial Unix variants from Sun and HP - dominating the market and selling licenses. On the other, Linux, free and open source. And in between, the strategic move that is repeating itself today: open source as a weapon for those who are behind. Netscape, crushed by Internet Explorer, opened its browser code and gave birth to Mozilla. IBM, unable to win with its own Unix variants, invested a billion dollars in Linux - not out of charity, but to erode Sun and Microsoft’s margins. In industry jargon it is called commoditize your complement: if you cannot sell that piece of the stack yourself, make sure your competitor cannot sell it at a premium either.
Twenty years later we know how it ended. Linux won almost everywhere: servers, the cloud, supercomputers, and - via Android - the pockets of billions of people.
The lesson for businesses: read the license, not the tweet
There is, however, one difference from the Linux era, and it is not a small one. “Open” today is a word that covers very different situations:
- DeepSeek releases under MIT, a clean and permissive license.
- The Qwen 3.8 Max weights arrived as a reduced checkpoint (text only, without the multimodal capabilities of the API version) and with license terms different from the Apache 2.0 of the 27B model.
- GLM-5.3 at launch was “open-weight” on paper, with the weights arriving weeks later.
- And in any case, many benchmarks accompanying these releases are measured by the vendors themselves: independent verification comes later, and often puts the numbers in a different light.
Not to mention the issue that we Europeans care particularly about: where these models run and where your data ends up. A self-hosted open-weight model on European infrastructure is one thing; its API hosted elsewhere is quite another.
This is precisely why we built AIDeskPro the way we did: plurality of engines as a principle, not a slogan. The landscape changes every ten days - August 2026 proved it - and no company should commit to a single LLM, or chase every release reading licenses and changelogs. We select, test and integrate the best models, serve them from European endpoints with zero data retention, and transparently label those that are not European. You choose the right engine for each task, with the peace of mind of knowing where your data is.
The OS wars taught us that the winner is not the one who cheers the loudest: it is the one who knows how to use the right tool at the right time. With LLMs, it will be the same.