Gemory-26B-A4B-GGUF

DigitalBonsai/Gemory-26B-A4B-GGUF

Gemory-26B-A4B-GGUF is a text generation model from DigitalBonsai. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.

Downloads · 30 days

7% of all-time downloads

All-time downloads

998

Public

Repo size

131 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.gguf52.6 GB · 100%

At a glance

Task: Text Generation
License: apache-2.0
Access: Public
Created: Sep 6, 2026
Updated: Sep 6, 2026
SHA: d76c1637

Try a prompt

Base models

huihui-ai/Huihui-gemma-4-26B-A4B-it-qat-q4_0-unquantized-abliterated

Spaces

Model card

From the Hugging Face model README

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.]</code></pre>

</div> </div> </div> </details>  <details> <summary> <span class="sum-title">⌘ Field calibration</span> <span class="sum-chip">sampling</span> </summary> <div class="gs-fold"> <p> 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. </p> <div class="gs-sampler"> <div class="gs-param"> <span class="pv">0.2</span> <span class="pl">Temp</span> </div> <div class="gs-param"> <span class="pv">0.98</span> <span class="pl">Top-P</span> </div> <div class="gs-param"> <span class="pv">80</span> <span class="pl">Top-K</span> </div> <div class="gs-param"> <span class="pv">0.03</span> <span class="pl">Min-P</span> </div> <div class="gs-param"> <span class="pv">1.15</span> <span class="pl">Repeat Penalty</span> </div> </div> <div class="gs-quote"> During testing, we found <strong>BF16 performs better at higher temperatures (0.8-1). Q4 quants perform excellently even at temperature 0.2.</strong> 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. </div> </div> </details>  <details> <summary> <span class="sum-title">⚙ Open the case file</span> <span class="sum-chip">training</span> </summary> <div class="gs-fold"> <p> Gemory was trained as a <strong>language-only BF16 LoRA</strong> on the abliterated Gemma 4 26B-A4B QAT-lineage checkpoint. </p> <div class="gs-stats"> <div class="gs-stat"> <span class="v">26B</span> <span class="l">Total model class</span> </div> <div class="gs-stat"> <span class="v">~4B</span> <span class="l">Active per token</span> </div> <div class="gs-stat"> <span class="v">25.36M</span> <span class="l">Training tokens</span> </div> <div class="gs-stat"> <span class="v">4546</span> <span class="l">Training examples</span> </div> </div> <div class="gs-table-wrap"> <table class="gs-table"> <thead><tr><th>Item</th><th>Value</th></tr></thead> <tbody> <tr><td>Training base</td><td><strong>Huihui-gemma-4-26B-A4B-it-qat-q4_0-unquantized-abliterated</strong></td></tr> <tr><td>Architecture</td><td>Gemma 4 26B-A4B routed MoE; 30 text layers; 128 experts; top-8 routing</td></tr> <tr><td>Primary target</td><td>Creative conversation / imaginative prose / roleplay / NSFW / ERP</td></tr> <tr><td>Fine-tuning method</td><td><strong>LoRA</strong></td></tr> <tr><td>Ordinary attention + shared MLP LoRA</td><td><strong>rank 16 / alpha 32 / dropout 0</strong></td></tr> <tr><td>Routed expert LoRA</td><td><strong>rank 4 / alpha 8 / dropout 0</strong></td></tr> <tr><td>Training examples</td><td><strong>4,546</strong></td></tr> <tr><td>Validation examples</td><td><strong>264</strong></td></tr> <tr><td>Training tokens</td><td><strong>25,364,445</strong></td></tr> </tbody> </table> </div> </div> </details>  <details> <summary> <span class="sum-title">◈ What was in the dossier?</span> <span class="sum-chip">dataset</span> </summary> <div class="gs-fold"> <div class="gs-grid"> <div class="gs-card"> <h4>Real Human Writing (Deslopped!)</h4> <p>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.</p> </div> <div class="gs-card"> <h4>Deterministic One-Epoch Shuffle</h4> <p>The 4,546 training examples were consumed once using the Trainer's deterministic distributed shuffle.</p> </div> <div class="gs-card"> <h4>Held-Out Validation</h4> <p>The preserved 264-example validation split remained separate throughout training, giving 4,810 total examples across training and validation.</p> </div> <div class="gs-card"> <h4>Language-Only Adaptation</h4> <p>No image examples were used during training. The multimodal path was preserved by exclusion and hash verification rather than retrained.</p> </div> </div> </div> </details>  <details> <summary> <span class="sum-title">⬡ Which build should I deploy?</span> <span class="sum-chip">files</span> </summary> <div class="gs-fold"> <div class="gs-builds"> <div class="gs-build"> <div class="q">BF16</div> <span class="badge">maximum fidelity</span> <p>The verified merged source for users with the memory. Ships with the more open creative temperature preset.</p> <code>Gemory-26B-A4B-bf16.gguf</code> </div> <div class="gs-build recommended"> <div class="q">Q4_K_M</div> <span class="badge">recommended balance</span> <p>The sensible first download for most local users: dramatically smaller than BF16 while retaining a quality-focused mixed Q4 layout.</p> <code>Gemory-26B-A4B-Q4_K_M.gguf</code> </div> <div class="gs-build"> <div class="q">Q4_0</div> <span class="badge">compact / qat-lineage</span> <p>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.</p> <code>Gemory-26B-A4B-Q4_0.gguf</code> </div> <div class="gs-build"> <div class="q">MMProj - Optional</div> <span class="badge">multimodal sidecar</span> <p>Shared projector for compatible llama.cpp-family runtimes. The same projector is used with the BF16 and Q4 main-model variants.</p> <code>mmproj-Gemory-26B-A4B-bf16.gguf</code> </div> <div class="gs-build"> <div class="q">MTP - Optional Recommended</div> <span class="badge">optional drafter</span> <p>Recommended huihui upstream speculative-decoding sidecar. It was not further trained; support and speedup depend on the runtime. Download and place alongside model.</p> <code>mtp-ggml-model-bf16.gguf</code> </div> </div> <div class="gs-quote violet"> The current Q4_0 and Q4_K_M files were produced directly from the unchanged verified BF16 source <strong>without an importance matrix</strong>. Both retained the shared multimodal path and passed real runtime generation tests. </div> </div> </details>  <details> <summary> <span class="sum-title">◫ Can she see what she's walking into?</span> <span class="sum-chip">multimodal</span> </summary> <div class="gs-fold"> <p> Yes—with an important distinction. <strong>Gemma 4 26B-A4B is multimodal upstream, but Gemory's training was language-only.</strong> The vision tower and multimodal machinery were deliberately excluded from LoRA training and verified unchanged afterward. </p> <p> 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 <strong>text generation</strong>; Gemory does not claim improved visual understanding. </p> </div> </details>  <details> <summary> <span class="sum-title">⚠ When does the cover slip?</span> <span class="sum-chip">limitations</span> </summary> <div class="gs-fold"> <ul> <li>This is a specialised creative-writing / roleplay fine-tune, not an attempt to improve every capability of Gemma 4.</li> <li><strong>A4B is genuinely sparse.</strong> 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.</li> <li>The underlying base is abliterated. Refusal and safety behaviour therefore differs from the original upstream instruction model.</li> <li>Creative outputs are sampler-sensitive. In this release, BF16 and Q4 intentionally ship with different temperature defaults; your frontend may override them.</li> <li>The LoRA trained only the language path. Multimodal behavior was preserved, not fine-tuned for improvement.</li> <li>The optional MTP drafter is upstream-derived and was not trained as part of Gemory's training.</li> <li>This model was not trained or evaluated as an autonomous agent. <strong>USE AGENTICALLY AT YOUR OWN RISK.</strong></li> </ul> </div> </details>  <details> <summary> <span class="sum-title">§ The Team Back At HQ</span> <span class="sum-chip">credits</span> </summary> <div class="gs-fold"> <ul> <li><strong>LampLighter</strong> — training guidance / tooling.</li> <li><strong>Morrow</strong> — conversion / quantisation tooling.</li> <li><strong>UltimateIntent</strong> — dataset generation / curation and GPUs.</li> </ul> <div class="gs-callout"> <div class="label">License and redistribution housekeeping</div> <p> The <a href="https://ai.google.dev/gemma/docs/core/model_card_4">official Gemma 4 model card</a> and Huihui base identify the applicable model license as Apache 2.0. Preserve required notices and attribution when redistributing derived or merged weights. </p> </div> </div> </details>  <div class="gs-footer"> <div class="line">Gemory — <span>Imagine What Happens Next.</span></div> <div class="micro">Gemma 4 // narrative operative // 26B total // ~4B active</div> </div> </div>

Task: Text Generation
License: apache-2.0
Languages: en
Created: Sep 6, 2026
Updated: Sep 6, 2026