🧠 BRAIN COMPOSITION · v6 · LIVE

WEVIA Brain v6
Souverain Pur

Module IA souveraine 100% local + cascade 21 providers · Ollama uncensored Dolphin-Llama3:8B en frontline · Zéro lock-in vendor

21 PROVIDERS 20 LIVE 100% RÉSILIENCE 7 MODELS LOCAUX 🎤 STT WHISPER 🌍 S151 OVH FREE 🔤 TRANSLATE LIVE 🚀 ABLITERATED PULLED 🔤 COHERE DIRECT 🔥 5 KEEPALIVE 🛠️ DIRECT API
7
Modèles Locaux
2 uncen + 1 code + 3 chat + 1 emb
8B
Params Dolphin
Q4_0 GGUF
8.0GB
Storage Local
7 modèles Ollama
21
Cascade Providers
Auto-fallback
0€
Coût IA / mois
Souverain pur
UNCEN
Dolphin Mode
Zero censorship

01🏠 Module IA Souveraine Pure (Local)

Modèles Ollama installés sur infrastructure WEVAL. Zéro appel externe. Maîtrise totale données + privacy GDPR/CNIL Maroc.

dolphin-llama3:8b
WevalBrain Frontline
⚡ UNCENSORED
Dolphin = Llama3 abliterated — variant uncensored officiel par cognitivecomputations. Aucun refus, aucun guardrail "as an AI assistant", parfait pour pharma B2B + research.
FamilleLlama 3
Params8 milliards
QuantizationQ4_0 (GGUF)
Taille4.66 GB
Latence~2-5s/req
ModeUncensored
all-minilm:latest
Embedding Engine
📊 EMBEDDING
Sentence embeddings 384-dim pour RAG WEVIA, KB search semantic, similarité documents pharma. Permet retrieval offline complete.
FamilleMiniLM-L6
Dimensions384
QuantizationFP16
Taille46 MB
Latence~50ms
UsageRAG + KB

02⚡ Cascade Cloud Souverain (Auto-fallback)

21 providers en cascade ordonnée. Si #1 fail → #2 → ... → #21. Zéro single point of failure. 20 LIVE actuellement, 100% résilience.

01
cerebras (fast)
Cloud Free
LIVE
02
groq
Cloud Free
LIVE
03
sambanova
Cloud Free
LIVE
04
mistral
EU Sovereign
LIVE
05
deepseek (V3)
Cloud Free
LIVE
06
nvidia-nim
Cloud Free
LIVE
07
gemini
Cloud Free
LIVE
08
alibaba-qwen
Cloud Free
LIVE
09
kimi-k2
Cloud Free
LIVE
10
openrouter
Cloud Free
LIVE
11
hf-router
Cloud Free
LIVE
12
github-models
Cloud Free
LIVE
13
together
Cloud Free
LIVE
14
cohere
Cloud Free
CFG NEEDED
15
replicate
Cloud Free
LIVE
16
zhipu (GLM-5)
Cloud Free
LIVE
17
cloudflare-ai
Edge Free
LIVE
18
groq-oss
Cloud Free
LIVE
19
ollama-s204 (Dolphin)
Local Sovereign
LIVE
20
ollama-s151
Local Sovereign
LIVE
21
ollama-qwen3
Local Sovereign
LIVE

03🔧 À Wirer (Manques Identifiés)

Modules qui pourraient enrichir le brain souverain pur. Wire-status: PENDING.

multilang-translate.php
✅ LIVE D627 · MarianMT replacement
🔤 TRANSLATE
Translation FR↔EN↔AR via Ollama tinydolphin S151 (1.3s) + HF Router CohereLabs Arabic fallback. Souverain pur 0€. Test FR→EN 1.3s ✅.
Use caseFR↔AR↔EN
Endpoint/api/multilang-translate.php
LocalOllama S151
Latence1.3s
Cloud fallbackHF CohereLabs Arabic
StatusLIVE
huihui_ai/qwen2.5-abliterate:7b
✅ PULLED D629 · S204 (S151 dolphin frontline)
🚀 ABLITERATED
huihui_ai/qwen2.5-abliterate:7b PULLED on S204 4.5GB. Honnete D4: cold start parfois 503 due to S204 RAM contention. Frontline reste S151 dolphin-llama3:8b. Real abliterated 70B en attente Scaleway 6mai.
ServerS204 (4.5GB)
Params7.6B
FamilleQwen2.5
QuantizationQ4_K_M
NoteCold start 503
StatusPULLED
brain-composition.php
API endpoint live
🔧 À WIRER
Endpoint /api/brain-composition.php — expose JSON live de la composition brain (modèles Ollama loaded + sovereign cascade + memory + GPU). Source-of-truth pour cette UI.
Endpoint/api/brain-comp.php
FormatJSON
Cache30s
SourceOllama API
Wire viaFWS write-safe
StatusWIRING
qwen2.5-coder:1.5b
✅ WIRED D624 · Code generation local
🛠️ CODE
Qwen2.5 Coder 1.5B — modèle spécialisé code Python/PHP/JS pulled D624. Variant 1.5B optimisé S204 sans GPU. Souverain pur 0.92GB. Remplace GitHub Models cascade.
FamilleQwen2.5
Params1.5 milliards
SpecialtyCode
Taille0.92 GB
Latence~2-3s
StatusLIVE
Cohere Command-R direct
✅ LIVE D629 · Arabic FR EN 731ms
🔤 COHERE
Cohere v2 API direct (HF Router credits depleted bypassed). Test Arabic "Hello doctor" → "مرحبا دكتور..." 731ms. 1000 calls/mois free. RAG-optimized + 13 langues natifs incluant Arabe Maghreb.
ProviderCohere Direct
Modelcommand-r-08-2024
SpecialtyArabe Maghreb + RAG
Latence731ms
Endpoint/api/cohere-chat.php
StatusLIVE
whisper-tiny
Speech-to-Text local
🔧 À WIRER
Whisper Tiny multilingual — STT offline FR/AR/EN ~75MB. Permet transcription appels HCP, voicemails, vocal commands WEVIA. Souverain audio.
FamilleWhisper
Varianttiny multi
Taille~75 MB
Langues99
Wire viawhisper.cpp
StatusNOT_INSTALLED