I ran a 9B model trained on Claude's reasoning on my 8GB GPU, and it made me rethink paying for Claude

Qwythos 9B v2, developed by Empero AI, is described as a "reasoning model," meaning it processes problems in a distinct step before generating a response. Based on Alibaba's Qwen3.5 9B, the model was trained using over 500 million tokens from Anthropic's Claude Mythos and Fable conversations. This training aimed to impart a similar problem-solving "style," with Empero's own "rethink" tool generating the step-by-step reasoning data. Empero AI targets the model for math, science, medical explanations, and cybersecurity, noting it is deliberately uncensored.

The model reportedly operates efficiently on an 8GB GPU, such as an RTX 3070, achieving speeds of 20 to 22 tokens per second with answers often appearing in under a second. This performance is attributed to the Qwen 3.5 architecture, which manages conversation memory by only doing so in a quarter of its blocks, allowing for a 14k context length on the specified hardware. This capability suggests that for tasks focused on explanations or problem-solving, a local model could potentially handle workloads often routed to paid cloud-based services.

However, Qwythos 9B v2 has reported weaknesses. It performs "a bit worse" at coding than its base Qwen model and can "confidently get facts wrong without web search." Additionally, Empero AI has not disclosed how it obtained the Mythos data used for training. This lack of transparency introduces an element of uncertainty, particularly given Anthropic's public stance on preventing data harvesting from its models.