---
base_model:
- Naphula/G4-Runic-Oarfish-26B-A4B-v1.2
language: eng
library_name: transformers
license: apache-2.0
tags:
- mergekit
- merge
- gemma4
- moe_karcher
widget:
- text: G4-Runic-Oarfish-26B-A4B-v1.2
  output:
    url: >-
      https://cdn-uploads.huggingface.co/production/uploads/68e840caa318194c44ec2a04/lcB0sODb6Rdo_JIQo4uh4.png
---

Proxy errors were kinda high at 3bpw layers this time, I'm unsure if this and other quants below 4bpw will work well 0_0

Text Completions TOTALLY not recommended!!! They're braindead

3.10bpw, H8.

Model's Card: 
# 🐟 G4 Runic Oarfish 26B A4B v1.2

![Runic_Oarfish](https://cdn-uploads.huggingface.co/production/uploads/68e840caa318194c44ec2a04/lcB0sODb6Rdo_JIQo4uh4.png)

This is a creative RP merge which combines [Musica](https://huggingface.co/AuriAetherwiing/G4-26B-A4B-Musica-v1) with the full [LORA](https://huggingface.co/zerofata/G4-26B-SFT-6) of [MeroMero](https://huggingface.co/zerofata/G4-MeroMero-26B-A4B). v1.2 also adds [Darkhn/Gemma-4-26B-A4B-Animus-V14.1-FFT](https://huggingface.co/Darkhn/Gemma-4-26B-A4B-Animus-V14.1-FFT), another high quality RP finetune.

It uses a custom method `moe_karcher` which adapts the standard `karcher` method to support mixture of experts. A few changes were made to the script to support the new `Gemma4` architecture. Note there were some issues setting up the merge, so the vision mode might be disabled.

**Runic Oarfish has some refusals but can be jailbroken or ablated as needed.**

`moe_karcher` merge with 3 models. This model produces much different output than v1 or v1.1 upon being tested.

## An improvement over v1

There is still slop with the "not x, but y" prose, though it writes better otherwise. It talked about a lighthouse / cursed island instead of the clockmaker shop.

i think 1.1 isn't as good as the original, it has a lot more subtle refusal than v1, shorter replies, and more negative Gemini-like behavior. it seems that `moe_karcher` is better than `moe_slerp`. 

A magnitude scan reveals that MeroMero had the highest L2 norm, followed by Animus, then Musica. This means that MeroMero had the "strongest pull" on the karcher direction.

100 iterations is enough to produce about the same fidelity as 1000

The base model `gemma-4-26B-A4B-it` was still chosen to be excluded for this version, but it might be added for v1.3 

```yaml
architecture: Gemma4ForConditionalGeneration
merge_method: moe_karcher
# base_model: B:\26B\google--gemma-4-26B-A4B-it
models:
  - model: B:\26B\AuriAetherwiing--G4-26B-A4B-Musica-v1
  - model: B:\26B\ApocalypseParty--G4-26B-SFT-6 # zerofata/G4-MeroMero-26B-A4B
  - model: B:\26B\Darkhn--Gemma-4-26B-A4B-Animus-V14.1-FFT
parameters:
  max_iter: 100
  tol: 1.0e-9
  router_strategy: karcher  # Options: karcher, average, first, random_init
  blend_experts: true  # Blend corresponding experts (expert[0] + expert[0], etc.)
dtype: float32
out_dtype: bfloat16
tokenizer:
  source: union
# chat_template: auto
trust_remote_code: true
name: G4-Runic-Oarfish-26B-A4B-v1.2
```

See [v1](https://huggingface.co/Naphula/G4-Runic-Oarfish-26B-A4B-v1) for more details of how to merge Gemma 4 MoE models.