tardellirs/aac-board-generator-moe-636m-ptbr-soup-GGUF overview
aac board generator moe 636m ptbr soup — GGUF GGUF llama.cpp do aac board generator moe 636m ptbr soup — MoE upcycled do gemma 3 270m 8 experts/top 4, ~636M to…
Runs locally from ~439.0 MB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| aac-board-generator-moe-636m-ptbr-soup.Q4_K_M.gguf | GGUF | GGUF | 439.0 MB | Download |
| aac-board-generator-moe-636m-ptbr-soup.Q5_K_M.gguf | GGUF | GGUF | 473.4 MB | Download |
| aac-board-generator-moe-636m-ptbr-soup.Q8_0.gguf | GGUF | GGUF | 649.2 MB | Download |
| aac-board-generator-moe-636m-ptbr-soup.f16.gguf | GGUF | GGUF | 1.19 GB | Download |
Model Details
| Model ID | tardellirs/aac-board-generator-moe-636m-ptbr-soup-GGUF |
|---|---|
| Author | tardellirs |
| Pipeline | — |
| License | gemma |
| Base model | — |
| Last modified | 2026-07-22T02:19:28.000Z |
Model README
---
license: gemma
language: [pt]
tags: [caa, aac, pictogramas, moe, gemma3, gguf, llama.cpp, papuguinho, pt-br]
---
aac-board-generator-moe-636m-ptbr-soup — GGUF
GGUF (llama.cpp) do aac-board-generator-moe-636m-ptbr-soup — MoE upcycled do gemma-3-270m (8 experts/top-4, ~636M total / 283M ativo)
para pranchas de CAA/AAC pt-BR (Estúdio Papuguinho). Variante: soup. Qualidade (juiz 0–10): soup k=1 = 9.16 (770m denso=7.8, teacher=9.2).
Arquivos
.f16 (fonte), .Q8_0, .Q5_K_M, .Q4_K_M. Roda a ~100–170 tok/s no CPU.
Rodar — escolha o k (experts ativos) no runtime
llama-cli -m aac-board-generator-moe-636m-ptbr-soup.Q5_K_M.gguf --override-kv gemma3moe.expert_used_count=int:1 \
-p "quero uma prancha de escola" -n 120> Requer llama.cpp com a arquitetura gemma3moe (grafo gemma3 com build_moe_ffn).
Notas
Multi-função (prancha + comandos add/remover). Não aplica teto de 12 quando se pede >12. Dados 100% sintéticos.
Autor: Stekel, Tardelli R. C. · Licença: Gemma.
Run tardellirs/aac-board-generator-moe-636m-ptbr-soup-GGUF with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models