LucidVitality-9B

LucidVitality-9B is a little goer. It's a roleplay or creative focused merge of two Qwen3.5 9b variants for people with absolute potatoes, like myself.

It marries the improved prose of Darkhn's Qwen3.5-9B-Animus-V13.0 with the lower looping, higher EOS exit, and slightly more coherency (compared to base) from Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED of DavidAU's making.

Background

I haven't merged anything for a long time, as it's been a hard time for finetuning. They rarely increase prose quality, often deeply lose intelligence over base (even when tuned for intelligence or agentic). Base model's getting tough to beat. So it was a pleasant surprise to find two models that each actually improved base in their respective ways. Props to those two fine finetuners!

Expected Performance

End result here, should be, roughly: more coherent, better instruct following, with a splash of Animus' improved prose.

Note: I did notice swearing at this model during instructions is not helpful (polite seems better). It's "specific form of instruct" sensitive (due to the roleplay content portion of the merge, I imagine). So if you do find it's not doing as you ask, try tweaking the prompt or prompt placing. It should be better behaved than uncensored base, if you prompt it right.

Seems to work, better than uncensored base at least, with ReCast in SillyTavern, with my overly complex post-processing prompts, returning usable output from a 2 stage post processing instruction about 4/5ths of the time (albiet sometimes a little sloppy on the editting, it is a 9b) compared to what is probably 1/5 or 0 with uncensored base, and also returns expressions catergorizations most of the time, using a relatively minimal about of total possible token output and proper EOS.

I've only tried it with fairly standard settings, and a limited number of prompts, so whether variable temperature, niche sampling or different prompt methods produce better results will be up to you/us to discover. Rough formulae for merge was similar to how I did SnowLotus:

Linear Gradient Merge > 50/50 dare ties back into the 'brains' model.

Files and Quantization

Included in the GGUF repo of the model is q4km, q5km variable quantizations, the fp16 gguf, and the imatrix output file (I used a slightly modified version of bartowski's calibration text, with a little more properly formatted prose injected: took 30 minutes on the potato!) so that others can easily make their own quants if wanted. None of the model files contain MTP because doing so with a uncensored, especially a finetuned model without tuning the MTP itself is a negative performance move and a bad general idea. Here: LucidVitality-9b-GGUF and more quants by mradermacher here

Merge Configuration

Linear Merge portion if anyone wants to replicate with other models, is as follows:

base_model: Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED

models:
  - model: Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED
    parameters:
      weight:
        - filter: self_attn
          value: [0.90, 0.60, 0.05]
        - filter: linear_attn
          value: [0.90, 0.60, 0.05]
        - filter: mlp
          value: [0.05, 0.60, 0.90]
        - filter: norm
          value: 0.5
        - value: 0.5

  - model: Qwen3.5-9B-Animus-V13.0
    parameters:
      weight:
        - filter: self_attn
          value: [0.10, 0.40, 0.95]
        - filter: linear_attn
          value: [0.10, 0.40, 0.95]
        - filter: mlp
          value: [0.95, 0.40, 0.10]
        - filter: norm
          value: 0.5
        - value: 0.5

merge_method: linear
dtype: bfloat16