---

license: gemma

language:

• pt

library_name: gguf

pipeline_tag: text-generation

base_model:

• tardellirs/aac-board-generator-moe-636m-ptbr

tags:

• aac
• augmentative-alternative-communication
• caa
• pictograms
• accessibility
• assistive-technology
• mixture-of-experts
• moe
• gemma3
• gguf
• llama-cpp
• quantized
• portuguese
• brazilian-portuguese
• on-device

---

AAC Board Generator · MoE 636M · pt-BR — GGUF

!License

!Language-009c3b.svg)

!Format

!Runtime

!Quant

!Speed

GGUF builds of aac-board-generator-moe-636m-ptbr —

a sovereign Mixture-of-Experts model that generates and edits AAC pictogram boards in Brazilian Portuguese,

running fully on CPU via llama.cpp. Built for Papuguinho.

≈636M total params · active params scale with k: ≈141M (k=1) · ≈212M (k=2, default) · ≈283M (k=3).

> ⚠️ Requires a llama.cpp build with the custom gemma3moe architecture. The stock upstream build does

> not yet include it.

---

Files

Two knobs, both baked into the file (no --override-kv needed):

k = active experts (1 = ≈2× faster · 2 = best quality) × quant (Q4 → Q8, smaller → higher fidelity).

Weights are identical across k; only runtime routing differs.

| Quant | Size | k=1 · fast (≈210 tok/s) | k=2 · best (≈90–130 tok/s) |

|---|---|---|---|

| Q4_K_M | 460 MB | aac-moe-636m-v2-k1-Q4_K_M.gguf | aac-moe-636m-v2-k2-Q4_K_M.gguf |

| Q5_K_M | 496 MB | aac-moe-636m-v2-k1-Q5_K_M.gguf | aac-moe-636m-v2-k2-Q5_K_M.gguf |

| Q6_K | 632 MB | aac-moe-636m-v2-k1-Q6_K.gguf | aac-moe-636m-v2-k2-Q6_K.gguf |

| Q8_0 | 681 MB | aac-moe-636m-v2-k1-Q8_0.gguf | aac-moe-636m-v2-k2-Q8_0.gguf |

| f16 | 1.2 GB | aac-moe-636m-v2-k1-f16.gguf | aac-moe-636m-v2-k2-f16.gguf |

Which file should I pick?

• 🏃 Smallest / fastest: k1-Q4_K_M (460 MB) — quality verified OK on both board generation and edit commands.
• 🏆 Best quality: k2-Q8_0 (681 MB).
• ⚖️ Balanced default: k1-Q5_K_M or k2-Q5_K_M.

Usage

# Board generation
llama-cli -m aac-moe-636m-v2-k1-Q4_K_M.gguf --jinja \
  -sys "$(cat corpo_chatbot.txt)" \
  -p "quero uma prancha de escola" -n 200 -st

# Ask for a specific count
#   -p "gera 5 pictogramas de praia"

# Board editing (agent) — pass the current board + a command, use corpo_comando.txt as system prompt

Each board line is label|category|search_terms. Edit commands return a compact DSL

(R| confirm, +| add, -| remove, T| retitle, C| columns, X| new theme).

Performance & Quality

• Speed: up to ≈210 tok/s (k=1) / ≈90–130 tok/s (k=2) on Apple-Silicon CPU/Metal; scales to server CPUs.
• Quality (LLM judge, 0–10): ≈ 9.1 (1B dense baseline = 7.8, teacher ≈ 9.2).
• Edit-DSL command following stays reliable down to Q4_K_M.

About

Model card, training data, evaluation and limitations: see the

base model repo.

100% synthetic data. Author: Stekel, Tardelli R. C. · Project: Papuguinho ·

License: Gemma.