Boulesis-v2.1-26B-A4B
Composite Gemma 4 RP model (QK task arithmetic + fused LoRA). Minor adjustment v2.
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.
[!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
What changed from v1 and v2
| v1 | v2 | v2.1 | |
|---|---|---|---|
| QK donor | Pantheon-Reasoning-1.1 | Pantheon-Reasoning-1.1-V2 | Same |
| QK alpha | 0.6 sliding q/k, 0.6 global q, 0 global k | 0.85 sliding q/k, 0.65 global q/k | Same |
| LoRA layers | all 30 | 10–28 | Same |
| LoRA targets | 55 | 35 | Same |
| Bake scale | 0.26 | 0.70 | 0.40 |
| Training data | mixed EN/RU | English only | Same |
MoE experts, the router, embeddings, MLP and the vision tower are identical to the abliterated body, verified by tensor comparison after every step.
How big the QK edit actually isI measured the task vector before merging instead of picking alpha by feel. Relative Frobenius norm against the base weights:
| Group | mean | max |
|---|---|---|
| sliding_q | 0.00278 | 0.00518 |
| sliding_k | 0.00263 | 0.00474 |
| global_q | 0.00356 | 0.00436 |
| global_k | 0.00286 | 0.00395 |
Per-row rotation at the release alphas, in degrees:
| Group | alpha | mean | max |
|---|---|---|---|
| sliding_q | 0.85 | 0.12 | 3.08 |
| sliding_k | 0.85 | 0.12 | 5.87 |
| global_q | 0.65 | 0.12 | 1.72 |
| global_k | 0.65 | 0.09 | 9.53 |
It's not much, but it's honest work.
Using it
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:
Feedback
This is an experimental merge, and I mostly test it on my own cards, so my coverage is narrow. If you run v2.1, I would really like to know:
- Whether you ran it with thinking enabled or disabled.
- How it holds up in long sessions — scene logic, character positions, who is addressing whom. This was the main complaint about v1 and the reason v2 exists.
- 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. If anyone wants to be involved in v3, I am open to that as well!
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!