0x alpha is a stealth OpenRouter model live since Aug. 20, free to use at 100 trillion tokens with a 1M-token context window, and its launch was the most viral on OpenRouter, reportedly doubling DeepSeek Flash usage. The core claim is that this is a Jipu/GLM-linked flash model: benchmarks, tokenizer behavior, and the Taiwan prompt test all point to a GLM-like lineage, while Bloomberg was said to have posted something “everything but confirmed” tying it to z.ai. The video splits on what matters next: if the daily gains are just new checkpoints, then there is no real continual learning; if the model truly learns over time, it could be a cheap, open-weight, self-improving challenger that enterprises and consumers may prefer over closed American labs.
0x alpha has been live since Aug. 20 and offered to the public through OpenRouter as a stealth launch, with the name hidden at first.
The launch offered 100 trillion free tokens and a 1M-token context window, with inference cost estimated at $15M-$20M a day.
Day-one tests were far weaker than day-six tests, and the model was described as improving day by day across the first week.
OpenRouter said the stealth model became its most viral launch, the new #1 model, and more than doubled DeepSeek Flash’s usage at launch.
The model was benchmarked at 80% on the Deep SWE subset, versus 65% for Fable 5 and 62% for GLM 5.3, but that subset was described as only one coding slice, not a frontier-level verdict.
A later “real usage” readout put the model around 63%, not 80%, while keeping a high intelligence score with reduced token output.
Tokenizer behavior and a Taiwan prompt test were both described as essentially identical to GLM, pushing the inference that 0x alpha is a Jipu-linked model.
As of the morning of the video, GLM was described as a new version of Jipu, and 0x alpha as a new version of Jipu; Bloomberg was said to have posted something that was “everything but confirmed” tying it to z.ai.
The model was described as a flash model: a smaller derivative of a much larger Jipu model, with rumored frontier models like “mythos” or GPT 5.6 at 6-10T parameters and the flash version at 2-3T if the rumor is right.
The model was described as likely costing only 30-50 cents per usage and as open-weighted in its entirety.
Elon Musk was cited as saying Chinese or open-source models would become mythos-level with cybersecurity risk by Q1 2027; Jipu founder Jitung replied it would take “two to three months,” not that long.
The day-over-day gains were framed skeptically as new checkpoints rather than proof of continual learning, and the blind-test analogy said 0x alpha, mid-tier ChatGPT, and cloud models often look mostly the same on many tasks.
If continual learning were real, the model could avoid catastrophic memory loss, keep prior skills, run 24/7, and catch frontier models by learning like a human, which is why enterprises may prefer open-weights models that do not send data to American labs.
OpenRouter’s stated policy was that Jipu retains prompts and responses but does not use them for training; that claim was called unverifiable, and the warning was not to feed proprietary code or personal data into the million-token context window because China could be collecting those inputs.
China was described as subsidizing or state-funding the free giveaway, distilling attacks on many model labs, and flooding the market with cheap intelligence to keep Western labs from charging premium prices; Meta and Zuckerberg were said to be trying the same cost-down strategy, but less effectively.
The sponsor pitch was Ledger’s agent stack: agents propose, humans approve, and the Ledger signer enforces, with availability on Cloud Code, Code X, Cursor, and broadly wherever LLMs are used, plus an open-source release available today.