Odette — our first Voice Chorus. Three voices, one head.

Odette-26B-A4B-Voice

Voices worked. So we asked: could voices be merged? The rabbit hole goes deeper — too much choice, too much room to experiment. This is the first chorus we ever sang.

Odette blends three Gemma 4 26B A4B voices into a single portable tensor (~1.5 GB) via the Voice tool's DELLA grafting (voice chorus).

Two steps — any Gemma 4 26B A4B GGUF you already have

# 1. Get the voice tool (one-time): https://huggingface.co/Wiself/voice
python3 voice.py path   # optional, puts `voice` on your PATH

# 2. Cast it onto any Gemma 4 26B A4B GGUF you already have
voice cast ./gemma-4-26b-a4b-it-Q4_K_M.gguf voice.safetensors --out ./voiced/gemma-4-26b-a4b-odette.gguf

Run it:

llama serve -m ./voiced/gemma-4-26b-a4b-odette.gguf --jinja

One file out. No adapter at runtime, no LoRA, no second model.

And you can too — with the voices you already have. Update the voice tool and run voice chorus, follow the prompts:

voice chorus
# Choose pulse:    the voice underneath — everything syncs to it, nothing shows of it
# Choose support:  the voice that joins first — color, flavor, harmony (Enter to skip)
# Choose lead:     the voice on top — carries the melody, wins 90% of disagreements

The recipe

Role	Voice	Graft	What it brings
pulse	Dark Scarlett	base	Dark M→F desire — the ground everything stands on
support	MeroMero	0.8	Anime character play — swoon, banter, living cast
lead	Pantheon Reasoning 1.1-V2	0.9	Forward-planning brain — thinks before it speaks

Method: DELLA magnitude pruning + TIES-sum combine, seed 42, per-voice epsilon (0.1 / 0.05). The duet (pulse→support at 0.8) was auditioned before the lead landed at 0.9. Full chain in voice.json; voice info shows it.

In their own words

"take the data that Pantheon is built on, pair it with full thinking traces, and let the model reason its way through character work — weighing tone, planning narrative beats, considering how a character would actually respond before committing to a line." — Gryphe, Pantheon-Reasoning-26B-A4B-1.1-V2

"genuine forward planning (considering character psychology, tone, and narrative direction), not post-hoc explanation." — Gryphe, Pantheon-Reasoning-26B-A4B-1.1-V2

MeroMero's card lists its training grounds outright: zerofata/Instruct-Anime, zerofata/Gemini-3.1-Pro-GLM5-Characters, zerofata/Roleplay-Anime-Characters — anime instruct plus roleplay-anime characters, per zerofata/G4-MeroMero-26B-A4B. Dark Scarlett's promise is simpler: "Passion ignited. Boundaries dissolved." — per ReadyArt/Dark-Scarlett-v1.0-26B-A4B.

Pantheon's numbers, linked not repeated: 0.70 clichés/response (−52.1%), 261.8-word thinking traces, 12.0% ngram overlap with base — the author's own writing metrics.

What's inside

voice.safetensors — the lm_head.weight tensor, source dtype preserved, arch-derived shape [262144, 2816], ~1.5 GB
voice.json — metadata: method della, pulse/donor/chain, per-graft density/epsilon/seed, config hash

Confirm after download: voice info voice.safetensors → lm_head.weight · [262144, 2816] plus Method/Pulse/Donor/Chain/Config lines.

Compatibility

Notes

Sampler tips from the source cards: temp 1.0 across all three; Pantheon adds MinP 0.05 and no repetition penalty (it corrupts thinking traces); Scarlett suggests Top-P 0.92; thinking ON at medium effort.
Usage terms carry over: Scarlett's donors are 18+, personal use only, you accept full responsibility — see the source card before sharing voiced models. All three sources are Apache-2.0 under Gemma terms.
Casting quantizes only the head to Q8_0 (near-lossless); all other tensors byte-copied.

Credits

ReadyArt — Dark Scarlett (dataset: GECFDO; fine-tuning: FrenzyBiscuit). zerofata — MeroMero (anime instruct + roleplay-anime characters). Gryphe — Pantheon Reasoning (thinking traces, judge-validated). The Voice tool grafted them. Base model is Google's, under Gemma terms.

Wiself/Odette-26B-A4B-Voice