Kazeia-engine/dist/MODELS.md

10 lines
612 B
Markdown

# Modèles (tablette /data/local/tmp/kz-engine, ou external storage prod)
| Rôle | GGUF | Taille | Note |
|---|---|---|---|
| Speaker reco | Qwen3.5-9B-Q4_0 | 5.0GB | qualité>4B, decode 8.4 |
| Speaker léger | Qwen3.5-4B-Q4_0 | 2.4GB | decode 15, prefill55 |
| Thinker reco | Qwen3Guard-Gen-4B.Q4_K_M | 2.5GB | bullets, decode11 |
| Thinker+ | Qwen3Guard-Gen-8B.Q4_K_M | 4.5GB | analyse fine |
| Dense rapide | Qwen3-4B-Q4_0 | 2.2GB | prefill103, decode17 |
Cascade 9B+Guard4B=7.5GB OK (16GB). 30B-A3B/35B=OOM, exclus. Q4_0>K_M pour NPU (repack). Source: unsloth/*-GGUF, mradermacher Guard. tied embeds Q6_K.