This 9B AI Model Runs on a Laptop — and Thinks Like a Giant (Qwythos-9B)

A 9-billion-parameter model that fits on an 8 GB laptop GPU and reasons like something ten times its size. Qwythos-9B (full name: Qwythos-9B-Claude-Mythos-5) is the most impressive small model I've run this year — with one serious caveat at the end.

What it is

Built by Empero AI on a Qwen3.5-9B base, post-trained on 500M+ synthetic reasoning traces. The numbers against its own base model are absurd:

Plus a 1,048,576-token context (yes, 1M) via YaRN. Quantized to Q4_K_M it's a 5.6 GB file — full GPU offload on an RTX 4060 laptop, running at 30–37 tok/s.

The live tests

I ran it in LM Studio through real tasks, all on camera:

• Step-by-step reasoning — clean chains, no rambling
• Native tool calling — proper JSON tool invocations
• JSON extraction — structured output that parses
• Multilingual translation — held up across languages
• Coding — generated code passed 10/10 executed tests

For a 9B running on a laptop, that test sheet reads like a mid-tier cloud model.

Where it fails

Two things you must know:

1. It's uncensored — no built-in safety layer. You're the safety layer.
2. It fabricates confidently. I asked about a study that doesn't exist; it invented detailed findings — authors, numbers, conclusions — without blinking. Great power, but verify everything it tells you.

Run it yourself

# Ollama
ollama run hf.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF:Q4_K_M

Or LM Studio: import the Q4_K_M GGUF and load with full GPU offload. Model: huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF

Verdict

The gap between laptop models and cloud models keeps shrinking, and Qwythos-9B is the sharpest data point yet: giant-class reasoning scores, 1M context, tool calling, working code — in 5.6 GB. Treat its factual claims like a confident intern's, and it's the best local daily driver an 8 GB GPU can run right now.