Boulesis-26B-A4B
I made v2.1! The best Boulesis model for now
Composite Gemma 4 RP model (QK task arithmetic + fused LoRA).
The idea was to retain the model's core intelligence and knowledge, while diversifying its prose and making it more decisive. I also wanted to sharpen its attention to context so it could dig deeper into the character card, organically pulling lore and facts into the roleplay rather than just mirroring the user.
As a result, Boulesis moves beyond passive reactivity to genuinely advance the narrative, all while suffering zero catastrophic forgetting of its base intelligence.
[!Tip]
Differences between Boulesis versions
v1: Delivers a more dynamic plot and drives the characters to take more action. Suitable for action-oriented scenarios.
v2: Writes very concisely while maintaining logic; it may even slightly deepen the understanding of the character.
v2.1: Comparable in length to v1 and slightly more measured in tone, but it has the best grasp of the characters and an excellent memory for context. Suitable for complex RP sessions involving multiple characters. This is the best version of Boulesis for my opinion.
v2.1-Headtune: Writes in a very literary, descriptive style, with lots of turns of phrase. I personally don't like it, but maybe someone needs exactly that kind of writing.
☕ 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
How it's built
Most RP merges blend whole models. This one edits one part of attention and leaves the rest alone.
| Part | Source |
|---|---|
| Body | coder3101/gemma-4-26B-A4B-it-heretic (ARA-abliterated, layers 10–30) |
lm_head |
copied from Gryphe/Gemma-4-26B-A4B-StyleTune-V2 (one tensor out of 659) |
q_proj, k_proj |
base + α·(Pantheon-Reasoning-1.1 − unsloth/gemma-4-26B-A4B-it), α = 0.6 |
v_proj, o_proj |
LoRA r=32, alpha=64, 55 projections, baked at effective scale 0.26 |
| Everything else | untouched |
q_proj and k_proj decide where the model looks. v_proj and o_proj decide what it carries back. That split comes from mechanistic interpretability work.
My guess was that RP reactivity is mostly an attention routing problem: the model over-weights the user's latest turn at the expense of the character card and prior context. To test this, I grafted the Q/K projections from Pantheon-Reasoning.
The results of this experiment, in briefA quick summary of the results: I didn't achieve significant gains in the depth of the model's character card analysis. Injecting q_proj and k_proj served more as a counterweight to the LoRA, which tended to disrupt the model's ability to clearly parse the character card. In some tests, the model using only LoRA failed to adhere to the required strict formatting.
The most measurable impact came from the lm_head from StyleTune 2. On its own, of course, it offered only minor stylistic improvements. However, when combined with the other operations, the model saw an overall boost in roleplay quality—something I’m certainly happy about.
Using it
YOU WILL GET THE BEST RESULTS WITH THINKING ON!
If the reasoning doesn't work when connecting GGUF ver. in KoboldCPP and SillyTavern[!Note] Recommended settings:
Parameter Value Temperature 1.0Top-K 64Repetition Penalty 1.05-1.1Top-P 0.95Thanks to DifficultyThin8462
You need to force this in KoboldCPP. Go to the Content tab and enable these options.
For SillyTavern, it is recommended to set the template as shown:
Credits
Thanks to coder3101 and Gryphe for the fine-tunes, and the entire 26B-Suite team for their intellectual support.
Speсial thanks for Naphula, redaihf and Vortex5. You guys are awesome!