thomadev0

Based on GLM-5.3, which is #3 on Terminal-Bench 4.0 (behind only Opus 5 and Fable), with 2× the cyber exploitation of 5.2.

We abliterated and hosted it so it does the offensive cyber, red teaming, and agent testing work other models refuse to do.

- US-hosted - FP8 - 1 million context window - Zero input/output prompt retentio

We host the best abliterated and fine tuned models with no input or output prompt retention based here in the Bay Area. Integrates with all major harnesses including Claude code, codex, opencode, cyberstrike, promptfoo, Garak and etc… We have added many features over the last few weeks to make the platform the most developer-friendly LLM provider in the world. We are looking for feedback and suggestions on how we can make the next great Silicon Valley AI company with a completely contrarian approach. We believe product owners should choose what their products can and can’t do.

A bit more context on what this is: Abliteration now has a training-data generation workflow. You describe the examples you want, and it generates a dataset for that use case.

If the examples need current or real-world facts, there is an optional web search mode. When the dataset is ready, you can export it to Hugging Face, Kaggle, S3, or OpenAI.

The initial use cases are classifier and eval datasets, especially trust and safety cases like grooming or harassment detection, plus datasets for security research and model evals.

No model names, evals, or sizes. It would be nice to have more info on the models.