Model Configuration Reference

Complete reference for PAL MCP Server model stack. Updated January 2026.

Overview

PAL uses a minimal, focused model stack - 4 local Ollama models + 5 OpenRouter cloud models. Each model has a specific purpose with no redundancy.

Local Models (Ollama)

These run on your RTX 5090 (32GB VRAM) via Ollama:

Model	Aliases	VRAM	Use Case
qwen2.5-coder:32b	coder, code, qwen	~19GB	Best local coding, 92 languages, rivals GPT-4o
deepseek-r1:32b	deepseek, r1, reasoning, think	~20GB	Best local reasoning, approaches O3
qwen2.5:3b	quick, fast, small, 3b	~2GB	Ultra-fast for simple tasks
dolphin3:8b	dolphin, uncensored, unfiltered	~5GB	Uncensored, no safety filters

Usage Examples

# Coding tasks
pal coder "Write a Python async HTTP client"

# Deep reasoning
pal reasoning "Analyze the time complexity of this algorithm"

# Quick simple tasks
pal quick "What's 2+2?"

# Unrestricted tasks
pal uncensored "Explain how X works without caveats"

Installation

# Install all 4 core models (default)
.\setup-ollama.ps1

# Minimal: just coder + reasoning
.\setup-ollama.ps1 -MinimalModels

# Full: adds qwen3:32b and phi4:14b
.\setup-ollama.ps1 -AllModels

Cloud Models (OpenRouter)

These run via OpenRouter API for tasks requiring massive context, multimodal, or cloud capabilities:

Model	Aliases	Context	Use Case
meta-llama/llama-4-maverick	maverick, llama4, vision, multimodal, images	1M	Image analysis, multimodal, 12 languages
deepseek/deepseek-v3.2	deepseek-v3, v3.2, deepseek-cloud, workhorse	164K	GPT-5 class reasoning, bulk work
minimax/minimax-m2.1	minimax, m2.1, m2	196K	Coding/agentic workflows, 49.4% Multi-SWE-Bench
z-ai/glm-4.7	glm, glm4, glm-4.7	203K	Multi-step reasoning/execution
mistralai/devstral-2512:free	devstral, devstral2, mistral-code, mistral-free	262K	Agentic coding specialist, FREE
xiaomi/mimo-v2-flash:free	mimo, mimo-flash, xiaomi	262K	Multimodal vision, FREE

Pricing Reference

Model	Input	Output	Notes
Llama 4 Maverick	$0.15/M	$0.60/M	Cheapest multimodal, 1M context
DeepSeek V3.2	~$0.27/M	~$1.10/M	GPT-5 class workhorse
MiniMax M2.1	$0.28/M	$1.20/M	Best for agentic coding
GLM 4.7	$0.40/M	$1.50/M	Z.AI flagship reasoning
Devstral 2	FREE	FREE	123B dense, MIT license
MiMo V2 Flash	FREE	FREE	Xiaomi multimodal

Usage Examples

# Image analysis
pal vision "Analyze this architecture diagram" --image diagram.png

# Heavy cloud reasoning
pal workhorse "Refactor this 2000-line file"

# Free agentic coding
pal devstral "Build a REST API with error handling"

# Free multimodal
pal mimo "Describe what's in this screenshot" --image screen.png

# Multi-step reasoning
pal glm "Debug this complex async issue step by step"

Complete Alias Reference

Local (Ollama)

coder, code, qwen       → qwen2.5-coder:32b
deepseek, r1, reasoning, think → deepseek-r1:32b
quick, fast, small, 3b  → qwen2.5:3b
dolphin, uncensored, unfiltered → dolphin3:8b

Cloud (OpenRouter)

maverick, llama4, vision, multimodal, images → meta-llama/llama-4-maverick
deepseek-v3, v3.2, deepseek-cloud, workhorse → deepseek/deepseek-v3.2
minimax, m2.1, m2 → minimax/minimax-m2.1
glm, glm4, glm-4.7 → z-ai/glm-4.7
devstral, devstral2, mistral-code, mistral-free → mistralai/devstral-2512:free
mimo, mimo-flash, xiaomi → xiaomi/mimo-v2-flash:free

Routing Logic

PAL automatically routes based on task type:

Task Type	Model Used
Coding	qwen2.5-coder:32b (local) or mistralai/devstral-2512:free (cloud, FREE)
Reasoning	deepseek-r1:32b (local) or z-ai/glm-4.7 (cloud)
Quick tasks	qwen2.5:3b (local)
Image analysis	meta-llama/llama-4-maverick (cloud) or xiaomi/mimo-v2-flash:free (FREE)
Agentic tools	minimax/minimax-m2.1 (cloud)
Heavy lifting	deepseek/deepseek-v3.2 (cloud)

Configuration Files

Local Models: custom_models.json

{
  "models": [
    {
      "model_name": "qwen2.5-coder:32b-5090",
      "aliases": ["qwen-coder", "coder", "code", "qwen"],
      "intelligence_score": 18
    }
  ]
}

Cloud Models: openrouter_models.json

{
  "models": [
    {
      "model_name": "mistralai/devstral-2512:free",
      "aliases": ["devstral", "devstral2", "mistral-code", "mistral-free"],
      "context_window": 262144
    }
  ]
}

VRAM Management

With RTX 5090 (32GB), you can run:

• One large model: qwen2.5-coder:32b (~19GB) + context
• Two medium models: deepseek-r1:32b + qwen2.5:3b
• Small + large: Any 32B model + qwen2.5:3b (runs alongside)

Ollama uses LRU eviction - least recently used models are unloaded when VRAM fills.

# Configure max concurrent models
$env:OLLAMA_MAX_LOADED_MODELS = 2

Optimized Variants

After downloading base models, run the optimizer to create RTX 5090 variants:

.\optimize-ollama-5090.ps1

This creates -5090 suffixed models with optimal quantization and context settings.