Science & Tech · AI · 6 hrs ago
Qwythos 9B runs on an 8GB GPU and aims to imitate Claude's reasoning style
A tech writer tested Qwythos 9B, a free AI model from Empero AI, on a graphics card with 8GB of memory.
The model is built on Alibaba's Qwen3.5 9B, and Empero says it trained it on more than 500 million tokens of conversations involving Anthropic's Claude Mythos and Fable.
Empero's tool wrote the reasoning examples, so the model's Claude influence is described mainly as a style rather than copied step-by-step thinking.
Empero has not said how it obtained the Claude data, amid Anthropic's public efforts to stop other labs from using its answers to train models.
A compressed version fit in under 7GB, allowing it to run on the writer's 8GB RTX 3070 graphics card.
Empero targets the model at subjects including math, science, medical explanations and cybersecurity, but says its coding is weaker than the base Qwen model.
It can also state wrong facts confidently without web search, so the test raises questions about whether a local model can replace a Claude subscription, but does not establish that it can.
Qwythos 9B is a free reasoning model from Empero AI that the writer ran on an 8GB graphics card.
The model is based on Qwen3.5 9B and was trained on more than 500 million tokens of Claude Mythos and Claude Fable conversations.
Empero says the model is aimed at math, science, medical explanations and cybersecurity.
Its reasoning steps were generated by Empero’s own tool, rethink, so the article describes Claude’s influence as a style rather than copied reasoning.
Empero has not said how it obtained the Claude data, and the model can confidently give incorrect facts without web search.
- Who
- Empero AI developed Qwythos 9B. The XDA writer tested it.
- What
- A free reasoning model trained to imitate aspects of Claude models’ reasoning style.
- When
- The article was published on October 24, 2026. The exact release date of Qwythos v2 is not stated.
- Where
- The writer ran it locally on an RTX 3070 graphics card with 8GB of VRAM.
- Why
- Empero AI aims the model at math, science, medical explanations and cybersecurity. The writer tested whether it could cover some uses of a Claude subscription.
This story does not have two clearly opposing sides.
Alibaba’s Qwen team released Qwen3.5 9B, the model Qwythos is built on.
XDA published the writer’s account of running Qwythos 9B v2 on an 8GB GPU.
- Model
- Qwythos 9B v2
- Base model
- Qwen3.5 9B
- Training data
- More than 500 million tokens of Claude Mythos and Claude Fable conversations
- GPU used
- RTX 3070 with 8GB of VRAM
- Quantization
- Q4_K_M, just under 7GB in size
- Publisher date
- October 24, 2026
What people are saying
coding is a bit worse than the base Qwen model
it can confidently get facts wrong without web search