---
license: apache-2.0
language:
- en
- zh
base_model:
- Qwen/Qwen3-4B
tags:
- qwen
- jinja
- dense
- TIES
- qlora
- unsloth
- uncensored
---

### better late than never?

> i made the base model a few months ago and never got around to finishing it.
> initially i did post train it properly... _but_
> i switched to a subsection of dolphin-R1 for Coral 1.6 and 2.0, which seemed like a way better idea than mixing openthoughts and openhermes, to try and buff out dynamic hybrid reasoning since it wasnt actually intentional.
> only after i finished the training run on the base model did i realize that _my entire fine tuning workflow didnt support thinking blocks_. that explained a lot.
> i didn't feel like fixing it then so i kept the base model and training set for later to finish and here we are.

to the like 12 people who downloaded the other Coral models, first of all thanks, i did not expect a single download, actual feedback on the models would be appreciated.

## a whole 4 billion paramaters!?

yes i know, what a _goliath_ of a model (at least for the Coral family so far), but seriously it's nothing special

- a TIES merge from several fine tunes of Qwen 3 4b.
  - including the original _and_ later updated checkpoint.
- then a QLORA-merge trained on a subsection of dolphin-r1 to cement the CoT.

# actual performance

- it passed rigourorous vibe testing, given you use the correct infrerence/chat settings.
  - keeps qwen 3 4b's chat settings.
- no refusal, TIES has a habit of removing post training censorship.
- unfortunatly, thinking got steamrolled (TIES again). it still outputs \<think> formatting but theyre always empty.
- normal 4b model limitations.

# quant guide

- bf16 / dont bother, just use;
- Q5 / near lossless and fits in a lot of devices