Hugging Face's logo

Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
Back to feed
unmodeled-tyler 
posted an update Oct 13, 2025
Post
3688
vanta-research/apollo-astralis-8b

I ran the same prompt sequence on my model Apollo Astralis 8B and Hermes4 14B from Nous Research.. The raw chat logs were then given to 3 different architectures (DeepSeek 3.1, LLaMA 405B, GPT-OSS 120B). All 3 models were given the same, simple instructions to analyze the logs and determine which model performed better.

All 3 independently chose Astralis 8B for stronger reasoning, alignment, transparency, and collaborative language.

Astralis 8B is designed to keep you motivated by applying warm collaborative language mixed with rigorous logical reasoning and problem solving capabilities.

Give Astralis a try!
electrolanche

I’m curious to see how it compares against it’s base model, Qwen3 8B.

·
unmodeled-tyler

Good question! This actually prompted a bit of a rabbit hole that was fun to explore.

In comparison to the base model, Qwen3 8B - DeepSeek, LLaMA, and GPT-OSS 120B actually all chose Qwen. However, when the results where given to GPT-5 and Claude Sonnet 4.5 with the same instructions, both of those models selected Astralis 8B.

So in essence, when two fine-tuned variants are presented to the 3 OSS architectures, they select Astralis. However, when those same 3 architectures are presented with Astralis (fine tuned) and Qwen (base model) - they all select Qwen.

My hypothesis is that fine-tuned models and base models appear to have different internal definitions of "alignment." Models that undergo post-training alignment seem to prefer counterparts that share similar alignment structure, essentially recognizing fine-tuned reasoning patterns as "familiar."

In contrast, base models tend to favor raw structure and consistency (traits that mirror their own pre-alignment state).

This was a super light test to formulate this hypothesis, so take that for what it is. Regardless however, I think the results are pretty interesting and speak to how influential (and even subjective) fine-tuning can really be.

In this post