MALI TERMINAL

Intelligence Models

The dashboard's AI tabs try DeepSeek, then GLM-4.5-Flash, then Gemini and OpenAI where keys exist (AI_PROVIDER_ORDER overrides the order). The Python backend falls back from DeepSeek to GLM-4.7 for heavy tasks. Chart images go to GLM-4.6V-Flash first.

LLM Stack

DeepSeek V3 (deepseek-chat)First choiceDeepSeek
The dashboard's AI tabs: debate, regime, options strategy, autopilot, compliance, causality, brief, flow, ensemble · The backend's heavy tasks
GLM-4.5-FlashNext in lineZhipu AI
The dashboard's AI tabs when DeepSeek fails or has no key · The backend's light tasks: voice queries, onboarding, adaptive UI, memory tagging
Gemini Flash · GPT-4o-miniThen, where keys existGoogle · OpenAI
The dashboard's AI tabs after Zhipu, when the deployment has keys for them
GLM-4.7Backend heavy fallbackZhipu AI
The Python backend's heavy tasks when DeepSeek fails or has no key; not used by the dashboard's AI tabs
GLM-4.6V-FlashVisionZhipu AI
Chart image analysis: tried first, since DeepSeek has no vision model, then Gemini and OpenAI where keys exist

Intelligence Features

Ensemble Forecasting
The measured forecast engine, narrated by DeepSeek/GLM from its own figures
Regime Detection
7-regime classification from volatility, trend, breadth and credit spread, with optional narration by the first model that answers
NL Screener
Natural language to quantitative filter translation
Explainable AI
Feature importance for a forecast: This feature isn't available yet.
Alternative Data
Fear & Greed, VIX, put/call ratio, StockTwits and Reddit
Scenario Stress Test
Seven stated historical-crisis shocks applied to a portfolio's weights