{"slug": "century", "title": "Century", "summary": "In August, Chinese AI labs released a series of frontier and near-frontier open-weights models, including DeepSeek-v4-Flash-0731, Alibaba's Qwen3.8-27B, Z.ai's GLM 5.3 and GLM 5.3 Flash, and Qwen3.8-Flash-Next, with the latter introducing an n-gram-based approach to add knowledge parameters with minimal compute cost. None of these innovations came from American companies, while Meta's Muse Glimmer scored 35 on the Intelligence Index, far below the frontier, and Anthropic and other U.S. labs focused on IPO preparations.", "body_md": "# Century\n\n\"There are decades where nothing happens, and there are weeks where decades happen.\"\n\nJohn Allsopp and I quote that line at each other a lot these days. Well, we did - until this month. August wasn't a week where decades happened. It was a whole century.\n\nThe month launched with the release of DeepSeek-v4-Flash-0731, a near-frontier model small enough and smart enough to herald the '[Business Watershed](https://thewatershed.markpesce.com/four-watersheds/)': for a few tens of thousands of dollars of kit, an SME could have an army of agents available for tasks.\n\nIn any other month, that would have been enough. But that was just the entrée.\n\nTwo weeks later, Alibaba released Qwen3.8-27B, a model [small enough](https://thewatershed.markpesce.com/does-the-home-watershed-arrive-tomorrow/) to fit on a well-equipped personal computer. It seemed very good. Almost impossibly good. The \"official\" benchmarking bore out the impression: an [Intelligence Index of 52](https://thewatershed.markpesce.com/52-70-byoai/) - by some margin the most \"intelligent\" small model ever released, and the rough equivalent of the best money could buy just six months earlier.\n\nThat landed the \"Home Watershed\", following in such close succession to the Business Watershed that it'll be forever impossible to sort out the difference between them. In the age of WFH/WFA, that feels wholly appropriate.\n\nJust these two would have been enough. But [Z.ai announced GLM 5.3](https://thewatershed.markpesce.com/a-crowded-frontier/), with a true frontier Intelligence Index of 60, *and* open weights. Pretty much anyone with enough of a hardware budget can now run a frontier model on their own kit.\n\nA few days after that, a mystery: \"Ox Alpha\". Offered without charge via OpenRouter, the model quickly gained a reputation for its capability while the world speculated about who had made it.\n\nA week later the mask dropped away, revealing [GLM 5.3 Flash](https://thewatershed.markpesce.com/the-pit-crew/). With that, the already impressive (and less than month-old) DeepSeek-v4-Flash-0731 had been surpassed on intelligence (57 vs 52) and beaten on price-per-task.\n\nBut the best was saved for last. The same day GLM 5.3 Flash came out, we got Qwen3.8-Flash-Next. This wasn't really a continuation of the Qwen3 series - it was a preview of Qwen4.\n\nAnd what a preview.\n\nThe key innovation rests on an old idea made new: the \"n-gram\".\n\nModels get smarter mainly by adding parameters. But parameters normally live inside the math: every token you process gets multiplied through layers of weights, so more parameters = more compute = more GPU cost. Qwen asked: is there a way to add raw \"knowledge\" parameters that cost almost no compute per token?\n\nThat's GLM 5.3 describing the why of n-grams. Here's what it offers on the how...\n\nA simple analogy: a standard embedding is a dictionary of single words. The n-gram embedding is a massive phrasebook sitting on a shelf across the room - you can't hold the whole thing, but you always know exactly which page you'll need next, so you send someone to tear it out while you keep working. You get the benefit of the phrasebook without it slowing you down or cluttering your desk.\n\nWhat it means: **models that are smaller and faster and smarter**. Very much on trend for this month.\n\n*Update: On Saturday, Tencent **previewed Hy4**, their own latest and greatest model, which compares favourably to GPT 5.3. Likely another Chinese firm at the frontier.*\n\nWhat are we to make of all of this?\n\nOne point first and foremost: *none* of these innovations came from American companies. *All* of them came from Chinese AI labs. America's one notable open-weights release of the month - [Meta's Muse Glimmer](https://thewatershed.markpesce.com/gradually-then-suddenly/), scoring 35 - is a perfectly serviceable worker, and nowhere near the frontier.\n\nMeanwhile, America's two IPO-bound frontier labs spent August accelerating their listing efforts, with Anthropic positing a $30T addressable market. Not unlikely - but also not Anthropic's for the asking.\n\nMy gut tells me these sudden and overwhelming Chinese advances have wrecked valuations everywhere in the AI stack *above* the physical layer. The physical layer itself - semiconductors, GPUs, data centres - is fine; free frontier weights only increase demand for compute. The hyperscalers have their profits baked in. It's the model makers who spent August making an unexpected transition: from bearers of strange and unique gifts to commodity providers of cognition.\n\nThere is a market for very high quality cognition. But will the Americans be selling it to anyone except their own?\n\nA month ago, the question would have seemed inconceivable. A century later, things look different.\n\nTo close, here's what GLM 5.3 has to say about that opening quote:\n\nLenin almost certainly never said it.", "url": "https://wpnews.pro/news/century", "canonical_source": "https://thewatershed.markpesce.com/century/", "published_at": "2026-08-30 21:45:43+00:00", "updated_at": "2026-08-30 21:52:02.243537+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "ai-products"], "entities": ["DeepSeek", "Alibaba", "Z.ai", "Qwen", "GLM", "Meta", "Anthropic", "Tencent"], "alternates": {"html": "https://wpnews.pro/news/century", "markdown": "https://wpnews.pro/news/century.md", "text": "https://wpnews.pro/news/century.txt", "jsonld": "https://wpnews.pro/news/century.jsonld"}}