---
format: "aidr-story-markdown/v1"
id: "70bf1d27d2ea37bf5f1ab968cc0e80ebccb3b738d8f0b5c65225e89fc3c03f39"
canonical_url: "https://aidr.today/70bf1d27?lang=en"
title: "Post-training 4B model with RL to fly Tesla Roadster 21% better than GPT-6 Astra"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-30T15:27:16.000Z"
category: "Models"
topics: ["llm"]
source_urls: ["https://skypilot.ai/blog/flying-roadster-rl","https://news.ycombinator.com/item?id=50305279"]
summary: "Qwen3-4B was RL-trained to fly a simulated flying Roadster and beats DeepSeek, Muse, Beam, Nemotron and GPT-6 Astra in storms, even when they get unlimited time to think."
---

# Post\-training 4B model with RL to fly Tesla Roadster 21% better than GPT\-6 Astra

> [Open the canonical story](<https://aidr.today/70bf1d27?lang=en>)

**Published:** 2026-10-30T15:27:16.000Z
**Category:** Models
**Topics:** llm

## Summary

Qwen3\-4B was RL\-trained to fly a simulated flying Roadster and beats DeepSeek, Muse, Beam, Nemotron and GPT\-6 Astra in storms, even when they get unlimited time to think\.

## Sources

- [Story source](<https://skypilot.ai/blog/flying-roadster-rl>)
- [Discussion](<https://news.ycombinator.com/item?id=50305279>)