Kitchoon-V4-26B-A4B

Kitchoon-V3 with two LoRAs baked in: Gemory for livelier characters and my own Anthropos for less repetitive swipes.

The Kitchoon series is designed as an RP- and ERP-oriented model, focusing on character development and featuring a neutral writing style.

What was changed

	Kitchoon-V3	Kitchoon-V4
Attention (q/k/v/o_proj)	V3	+ Gemory, all layers, at 0.5 + Anthropos l07, layers 13–23, at 0.7
Dense MLP (gate/up/down_proj)	V3	+ Gemory, all layers, at 0.5 + Anthropos l07, layers 13–23, at 0.7
MoE experts	V3	+ Gemory at 0.5
lm_head, router, embeddings, vision tower	V3	identical

---

☕ Support

If this model was useful to you, you can support my work with a small donation.
It helps me pay for GPU time and keep releasing new models and datasets.

Gram (TON network): UQB6cmEHWIcuiwsjgO_C0Uz4fyX3Puj09BAB3hEVVpDDcQ7K
Boosty donation: Link

---

Details

• Gemory (UltimateIntent/Gemory-26B-A4B-LORA) — a LoRA on attention, dense MLP and MoE experts. In testing it made characters react more individually and talk to each other more.
• Anthropos l07 (SubMaroon-exp/Anthropos-LoRA-v2-EXP) — my decision-point LoRA. It teaches the model to open replies and paragraphs more like the base model does, and leaves the rest of the text to the IT model.
• Baked with stochastic bf16 rounding, so the baked strength matches the tested one. Verified in llama.cpp: V4 matches V3 + the merged LoRA.

Merge settings

LoRA	Strength	Parts	Rank
Gemory	0.5	attention + dense MLP, all layers	16
Gemory	0.5 × 0.25 (correction)	MoE experts, all layers	4
Anthropos l07	0.7	attention + dense MLP, layers 13–23	16

Using it

Recommended settings:

Parameter	Value
Temperature	0.85
Min-P	0.05
Reasoning budget	normal or short

Works with and without thinking. Without thinking is the cleanest. With thinking, prefer a normal or short reasoning budget: very long (auto) reasoning can occasionally produce odd words.

What to expect

In testing, compared to Kitchoon-V3:

• swipes open differently much more often — fewer replies starting with the same words;
• a different character sometimes reacts first, not always the same one;
• characters react more individually: teasing, pushing back, talking to each other, not only to you;
• with thinking, the reply follows the plan from the reasoning but is rewritten, not copied from the draft.

Known quirks:

• now and then an odd word: "own" or "quite" in the wrong place, or a doubled word ("her her"). Mostly in very long reasoning;
• swipes vary in how the scene is delivered more than in what happens — the main story beat still comes from the scene;
• smirk and the em dash are still around.

Feedback

This is an experiment, and I mostly test it on my own cards, so my coverage is narrow. If you try it, I would really like to know:

• How it holds up in long sessions — scene logic, who is where, who is talking to whom. I haven't tested this enough yet.
• How thinking behaves for you, and at which reasoning budget.
• Whether the odd words bother you in actual play.
• Your sampler settings, if you got clearly better or worse results than with the defaults.

Open a thread in the Community tab, I read all of them.

Credits

Thanks to UltimateIntent for Gemory, to Tetratype for the antislop base under Kitchoon-V3, to coder3101 for the heretic version it is built on, and to Sam Paech for the auto-antislop pipeline. Thanks to the entire 26B-Suite team for their intellectual support. Special thanks to Naphula and redaihf. You guys are awesome!

And special thanks, of course, to Indexnusrefather for his excellent dataset for original Kitchoon

HHS1E, Nikita Eshkere and all Kitch00n ON chat - Sending you my regards!