GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Husky110
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Husky110
Model Tree
ReadyArt/gemma-4-31B-it-scotoma-2
zerofata/G4-MeroMero-31B
TheDrummer/Artemis-31B-v1.1
Input Modalities
Output Modalities
Supported Functionality
A roleplay / creative-writing merge uniting three traits in one model:
Conceptually a rebuild of the Mero-Artemis (by sophosympatheia) recipe,
but laid on scotoma-2 as the carrier instead of bare gemma-4-it.
I was a fan of this model, but it had some issues with token-choices (I\\l do that. instead of I'll do that. for example) and generating more slop.
In theory, scotoma's cleaned-up token distribution wins the selection and should fight against common Gemma 4-slop.
Oh - and before you ask "Why didn't you use Artemis 1.2?" -> Caliperbench indicates fewer errors and more coherence in Artemis 1.1, compared to 1.2. The idea was to get the character-fidelity of MeroMero, the prose of Artemis and put that on top of a less sloppy base-carrier (scotoma-2).
Also - if you are looking for a roleplay-frontend to check out - come take a look at LettuceAI.
The ingredients stayed the same, but their distribution and impact changed. When testing 0.1 automatically, everything looked fine. But manual tests with LettuceAI showed some deep problems.
Words got glued together, Reasoning went into a spiral that took up all the budget and produced no output...
Doing some more manual tests (every ingredient with their i1_Q4_K_M or Q4_K_M version), basically throwing the same prompt, sampler and scenario with 3 user turns and 3 swipes each revealed that Artemis was the culprit here. I've been told that Artemis 1.2 fixed that issue, but brought other quirks with it. So I wanted to see if I can fix those found problems. And I think I did, by changing "What ingredient does what?" in this version. Version 0.1 basically mashed everything together, so that every model-ingredient is in every layer. This version tries to target specific tasks.
This is the current recipe:
ties + rescale: true.lambda 1.0, normalize: false, dtype/out_dtype: bfloat16.density: 1.0 on embed_tokens; embed / norm / lm_head
weight 0.25 for both ingredients, so scotoma's tokens stay dominant.mero-artemis-scotoma_v0.2.yml.This model is built for mature fiction and will produce explicit NSFW content. 18+ only. It is a fictional roleplay / creative-writing model with no factual reliability and no advisory or safety function. Users assume full responsibility for generated outputs; the author disclaims liability for misuse. Released under the Apache-2.0 license.
Huge thanks to the people whose work this merge stands on: