license: apache-2.0
base_model: Qwen/Qwen3-4b-thinking-2507
tags:
- reasoning
- code
- inference
- chat
- multimodal (if applicable)
library_name: transformers
datasets:
- nvidia/AceReason-1.1-SFT
Astral-4B-Preview
Astral-4B is a specialized reasoning-focused language model developed as part of the Astral series, designed to deliver high-fidelity, step-by-step reasoning with configurable depth. Built upon the Qwen3-4b-thinking-2507 foundation, this variant has been fine-tuned on the nvidia/AceReason-1.1-SFT dataset to enhance logical coherence, problem-solving capability, and structured thinking.
This model is currently in preview and intended for research, evaluation, and development use. Feedback is encouraged to guide future iterations.
Usage Instructions
To invoke the model correctly, include a reasoning-level indicator in the system prompt using the Reasoning-level: directive. The available levels are:
| Level | Behavior |
|---|---|
none |
No reasoning trace generated, direct response only. |
low |
Minimal internal reasoning |
medium |
Creates a reasoning trace thats not too long nor too short |
high |
Second highest reasoning depth |
ultra |
Maximum depth reasoning |
Note: The absence of a valid reasoning level will result in undefined behavior. Always specify one.
Example Prompt (ChatML Format):
<|im_start|>system
Reasoning-level: high
<|im_end|>
<|im_start|>user
What is the capital of France?
<|im_end|>
<|im_start|>assistant
<think>
Important Notes
- This is a preview release. Performance may vary across edge cases or non-standard inputs.
- For production applications, please wait for the official release.