YouTube Creator Felix Kjellberg's Local Model Faces Account Suspension, One Instance Clearly Labeled as Distillation
YouTube creator Felix Kjellberg released the local model Ajax and stated in a video that OpenAI suspended his account twice during development. The emails he showcased explicitly labeled one of the suspensions as distillation, which refers to using the output or inference of one model to train another model.
Ajax is a fine-tuning of Alibaba's Qwen3.5-9B, with 9 billion parameters, aimed at his self-hosted workspace Odysseus rather than the leaderboard. The project page states it is designed as a persistent agent to handle search, browsing, email, calendar, and to-do tasks, with conversations remaining on the user's own hardware. Refusal behaviors have been reduced to provide a less restricted experience; he still maintains boundaries against harmful requests and requires users to take responsibility. The weights were not publicly disclosed at launch, and hardware requirements were also not announced.
Odysseus is set to launch in May 2026, positioned as a free, open-source, self-hosted alternative to ChatGPT, licensed under AGPL-3.0 or later. Reports indicate that his development machine used 10 graphics processors, 8 of which were modified RTX 4090s with 48GB of memory; the end product targets home computers. The video went live around October 1.
Kjellberg mentioned he wanted to "distill a little" to improve Ajax and used the output of the OpenAI model as seed data. After the first suspension, he successfully appealed and restored his account; however, after running the model again to generate seed data, his account was suspended again. He questioned in the video, "How did they know?" The emails did not provide specific examples. OpenAI has not publicly commented on this account. Their user terms prohibit automated or programmatic extraction of outputs and forbid using outputs to develop competing models.
He described the asymmetry as: the other party works hard to take everyone's data to build AI machines, while he cannot use ChatGPT's output to train models. This part is his statement, not OpenAI's investigative conclusion. What stands out is not the 9 billion parameters themselves, but the suspension emails and this contrasting statement.
This is not about model procurement but about enforcing terms. The buyer is an audience looking to replace cloud subscriptions with home agents; the seller is OpenAI, which continues to charge the same group for ChatGPT. The incident is driven by account suspensions, not API price reductions. Beneficiaries include Qwen, which has already opened weights and can be fine-tuned, and Odysseus, which treats the local workspace as the default entry point; under pressure are individual developers who use outputs as seed data, and once distillation detection is publicly discussed, they will need to explain the detection boundaries of closed-source interfaces. Money did not change hands; what changed was the source of seed data.
Source: Public Information
ABAB AI Insight
Felix Kjellberg first released Odysseus in May 2026, integrating chat, agents, email, and calendar into a self-hosted workspace, then fine-tuned Ajax from Qwen3.5-9B, rather than pre-training from scratch. The path is similar to a batch of small models around 2025: open weights as a base, closed-source model outputs as seeds. The difference he publicly highlighted is the removal of refusal behaviors while including the first suspension email in the video. OpenAI has not responded to this account, but the terms clearly state that outputs cannot be used to develop competing models.
Capital is not present in this video. Alibaba opened the 9B weights, allowing individuals to fine-tune using modified graphics cards; OpenAI retains outputs in the account terms. Kjellberg's second suspension occurred after he successfully appealed and used the same interface to generate seed data again. The detection methods were not specified in the emails. Existing industry practices include interface logs, bulk feature extraction, and output watermarking; in 2025, OpenAI and Microsoft publicly stated they discovered individuals distilling inference models through interfaces, specifically pointing to training controversies like DeepSeek. The personal account suspension was based on the same terms, not a different set of laws.
In contrast, DeepSeek gained attention from distillation controversies, and the division of "open weights plus private alignment data" along the Mistral and Llama paths. Ajax is stalled in the latter half: the base is someone else's, the alignment data is alleged to come from closed-source outputs, and the weights have not yet been released. The industry position is not a competition of foundational models but a local agent aiming to replace cloud subscriptions with a one-time fine-tuning. Audience alignment is due to existing training corpus controversies, not because the 9B model surpassed closed-source frontiers.
Structurally, the transfer of pricing power is intercepted by terms. Open weights reduce the copying cost of model files to nearly zero, while closed-source companies retain executable control over outputs: viewable, usable, but not trainable against competitors. The mechanism involves accounts, logs, and suspension notifications, without needing to win copyright lawsuits first. Web data training and interface distillation are therefore not the same contractual relationship. The former deals with publicly available corpus without contracts, while the latter involves users who have agreed to terms.
ABAB News · Cognitive Laws
Weights can be given away for free, but outputs are still priced per use.
Winning an appeal once does not rewrite the terms.
Local models are free, but seed data is not.