Gemma 4 // 26B-A4B // Narrative Operative

Gemory

Fast. Precise. Dangerously imaginative.

A creative-writing and roleplay operative built to enter the scene, read the room, make the move, and keep going. Gemory carries 26B total parameters with roughly 4B active per token—a little femme fatale with very little interest in wasting motion.

Gemma 4 26B-A4B Routed MoE ~4B Active Creative Writing Roleplay Uncensored ERP ◆ GEMORY // FIELD FOOTAGE BRIEF → ENTER → READ THE ROOM → MAKE CONTACT → CHANGE THE SCENE → MOVE ON Mission Profile No wasted motion.

Gemory is tuned for imaginative conversation with intent: distinct voices, concrete detail, escalating pressure, and scenes that change because something happened in them. Gemory brings narrative presence, consequence, and forward motion.

Architecture 26B / A4B Disposition Uncensored Specialty Narrative Infiltration ◆ Mission briefing: who is Gemory? start here

Gemory is a creative-writing and roleplay operative designed to go deep undercover. Fast, smart and dangerously seductive, her embrace will leave you breathless.

A LoRA fine-tune of Gemma 4 26B-A4B trained from Huihui's abliterated QAT-lineage base, the A4B architecture gives Gemory a peculiar personality as a model: a large 26B total model body with only 4B parameters active for each token. That makes her extremely fast on even modest hardware, while staying true to character voice, dialogue, narrative movement, imaginative commitment, and conversational continuity.

◆ Deep Cover

Characters inhabit their roles instead of collapsing into one generic assistant voice. Motives, habits, friction, attraction, disagreement, and personality are allowed to survive the turn boundary.

◇ Close Contact

Dialogue favors rhythm, subtext, interruption, hesitation, specificity, intimacy, awkwardness, and character-shaped phrasing over polished-but-flat exposition.

◆ Objective Focus

Gemory prefers to enter the imagined space and do something with it: establish concrete detail, make a choice, create a consequence, and commit to a direction.

◇ Clean Extraction

Whether scenes burn slow or move fast, she handles them with grace. Escalate, reveal, complicate, redirect, settle, or leave a new problem, she's ready for the challenge.

✦ Operational profile: what kind of writing? the fun bit

Gemory is not a single-genre style model. The target is creative presence with continuity: prose that behaves as though it is already inside the situation, watching the same room, remembering the same people, and noticing when the balance of power changes.

Reads the Room

Atmosphere, sensory texture, implication, place, social pressure, and the feeling that the world extends beyond the current paragraph.

Makes Contact

Character-specific dialogue with interruption, humour, tension, intimacy, conflict, silence, awkwardness, and subtext.

Keeps Her Cover

Carry forward established details, motives, relationships, emotional consequences, physical state, and unresolved threads.

Precision Work

Prefer the peculiar image, concrete gesture, sharp line, or unexpected consequence over the safest stock phrase within reach.

Leaves a Mark

A scene should be different after the response than it was before it. The setup is material to act upon, not something to admire from a safe distance.

Any Theater

Designed for open-ended creative conversation rather than one fandom, genre, character archetype, relationship structure, or prose imitation.

⌁ How do I brief the operative? prompting

Give Gemory enough intel to know what world she is entering, who matters, and what pressure is already present. You do not need to hand her a screenplay. A compact premise with useful constraints is often enough.

You can provide a full story bible when continuity demands it. You can also give her one loaded situation and let the model do some of the imaginative labor. Brief the objective; no need to micromanage every footstep.

Example mission brief

Use as much or as little of this as the scene needs. It is a scaffold, not a required schema.

## CURRENT MISSION PARAMETERS
Premise / objective:
[What is happening? What wants to happen next?]
User role / perspective:
[Who the user controls or how they participate.]
Characters:
[Only the traits, motives, relationships, secrets, or constraints that matter.]
Setting:
[Place, period, atmosphere, relevant world rules.]
Tone:
[e.g. gothic romance, psychological drama, pulpy espionage, absurd comedy.]
Narrative pressure:
[What wants to change, collide, escalate, tempt, fail, or be discovered?]
Continuity constraints:
[Facts that must remain true.]
Opening frame:
[Where the scene begins and what is already in motion.]

⌘ Field calibration sampling

These are the recommended sampler settings for the Q4 variants (recommended quants for most users). They are included in the GGUF metadata so a user can load the model and get a sensible creative experience without immediately tuning a sampler panel.

0.2 Temp 0.98 Top-P 80 Top-K 0.03 Min-P 1.15 Repeat Penalty During testing, we found BF16 performs better at higher temperatures (0.8-1). Q4 quants perform excellently even at temperature 0.2. Tune for your taste and scenarios. Frontends can still override model metadata, so if your runtime imposes its own preset, reproduce these values manually before judging the model. ⚙ Open the case file training

Gemory was trained as a language-only BF16 LoRA on the abliterated Gemma 4 26B-A4B QAT-lineage checkpoint.

26B Total model class ~4B Active per token 25.36M Training tokens 4546 Training examples ◈ What was in the dossier? dataset

Real Human Writing (Deslopped!)

The training corpus contained only real pre-AI human dialogues, both sfw and nsfw, painstakingly deslopped and handcurated. Spelling, grammar, punctuation were all cleaned/corrected before training.

Deterministic One-Epoch Shuffle

The 4,546 training examples were consumed once using the Trainer's deterministic distributed shuffle.

Held-Out Validation

The preserved 264-example validation split remained separate throughout training, giving 4,810 total examples across training and validation.

Language-Only Adaptation

No image examples were used during training. The multimodal path was preserved by exclusion and hash verification rather than retrained.

⬡ Which build should I deploy? files BF16 maximum fidelity

The verified merged source for users with the memory. Ships with the more open creative temperature preset.

Gemory-26B-A4B-bf16.gguf Q4_K_M recommended balance

The sensible first download for most local users: dramatically smaller than BF16 while retaining a quality-focused mixed Q4 layout.

Gemory-26B-A4B-Q4_K_M.gguf Q4_0 compact / qat-lineage

A conventional Q4_0 derivative from the verified BF16 merge. The upstream base originated from a Q4_0 QAT pipeline; Gemory's training itself was not a new QAT stage.

Gemory-26B-A4B-Q4_0.gguf MMProj - Optional multimodal sidecar

Shared projector for compatible llama.cpp-family runtimes. The same projector is used with the BF16 and Q4 main-model variants.

mmproj-Gemory-26B-A4B-bf16.gguf MTP - Optional Recommended optional drafter

Recommended huihui upstream speculative-decoding sidecar. It was not further trained; support and speedup depend on the runtime. Download and place alongside model.

mtp-ggml-model-bf16.gguf The current Q4_0 and Q4_K_M files were produced directly from the unchanged verified BF16 source without an importance matrix. Both retained the shared multimodal path and passed real runtime generation tests. ◫ Can she see what she's walking into? multimodal

Yes—with an important distinction. Gemma 4 26B-A4B is multimodal upstream, but Gemory's training was language-only. The vision tower and multimodal machinery were deliberately excluded from LoRA training and verified unchanged afterward.

The GGUF package includes the shared projector, and tested llama.cpp/LM Studio paths retained text, vision, and video support. The creative-writing claims on this card refer specifically to text generation; Gemory does not claim improved visual understanding.

⚠ When does the cover slip? limitations

• This is a specialised creative-writing / roleplay fine-tune, not an attempt to improve every capability of Gemma 4.
• A4B is genuinely sparse. Only roughly four billion parameters are active per token. On difficult long-form narrative logic, Gemory can feel less deep or less deliberate than a much larger dense model even when the prose is perfectly serviceable.
• The underlying base is abliterated. Refusal and safety behaviour therefore differs from the original upstream instruction model.
• Creative outputs are sampler-sensitive. In this release, BF16 and Q4 intentionally ship with different temperature defaults; your frontend may override them.
• The LoRA trained only the language path. Multimodal behavior was preserved, not fine-tuned for improvement.
• The optional MTP drafter is upstream-derived and was not trained as part of Gemory's training.
• This model was not trained or evaluated as an autonomous agent. USE AGENTICALLY AT YOUR OWN RISK.

§ The Team Back At HQ credits

• LampLighter — training guidance / tooling.
• Morrow — conversion / quantisation tooling.
• UltimateIntent — dataset generation / curation and GPUs.

License and redistribution housekeeping

The official Gemma 4 model card and Huihui base identify the applicable model license as Apache 2.0. Preserve required notices and attribution when redistributing derived or merged weights.

Gemory — Imagine What Happens Next. Gemma 4 // narrative operative // 26B total // ~4B active