Astral 1.5 4B is a updated version of the Astral 4B model, trained for STEM tasks and agentic usage.
The model was trained on our Astral-1.5-Post-Training-SFT and Astral-1.5-Post-Training-CRLFT datasets.
Comparisons
When asked for a TailwindCSS HTML page for a photo editing software models gave the following (zero-shot responses):
Qwen3 4B 2507 Thinking (8101 tokens used)
Astral 1.5 4B (6088 tokens used)
GPT-5.2
Usage
You may utilize the following chat template to do regular chat completions with the model:
<|im_start|>user
What is the capital of France?
<|im_end|>
<|im_start|>assistant
<think>
CORRECT:
The
or to use it as an agent, you must prefill <think>Agent mode is on</think> like so:
<|im_start|>user
What is the capital of France?
<|im_end|>
<|im_start|>assistant
<think>
Agent mode is on
</think>
