Astral-1.5-4B-Preview-fixed

Astral 1.5 4B is the medium sized model in the Astral 1.5 family. It was fine-tuned from Qwen3 4B 2507 Instruct on LucidityAI/Astral-1.5-Post-Train-Dataset.

Note: This is on a different repo than the original preview, as for there was an issue merging the original model, leading it it being highly unreliable.

We reintroduce the reasoning effort selection as seen in Astral 1's preview model, yet not used in Astral 1.

The following modes are available for Astral 1.5:

• low: Generates a smaller reasoning trace
• medium: Generates a decent sized reasoning trace:
• high: Generates a generally unrestricted in size trace
• agent: Genreates no reasoning trace. Intended for agentic use.

Usage example:

<|im_start|>reasoning_level
medium
<|im_end|>
<|im_start|>user
What is the capital of France?
<|im_end|>
<|im_start|>