Model Selection & Tuning

Selecting the right LLM for the purpose, plus fine-tuning and quantization strategies

Model Selection Framework

    flowchart TD
    A["Requirements analysis"] --> B{"Data security"}
    B -->|"Internal data cannot leave"| C["Self-hosted<br/>open-source model"]
    B -->|"External API allowed"| D{"Task complexity"}
    D -->|"Simple classification/summarization"| E["Small model<br/>Claude Haiku / GPT-4o-mini"]
    D -->|"Complex reasoning"| F["Large model<br/>Claude Opus / GPT-4o"]
    D -->|"Domain-specific"| G["Consider fine-tuning"]

    style A fill:#1E3A5F,stroke:#1E3A5F,color:#fff
    style C fill:#EA580C,stroke:#C2410C,color:#fff
    style E fill:#16A34A,stroke:#15803D,color:#fff
    style F fill:#2563EB,stroke:#1D4ED8,color:#fff
    style G fill:#7C3AED,stroke:#6D28D9,color:#fff

Major LLM Comparison (2025)

For real-time, objective metrics on performance, cost, and latency, see the AI Model Benchmarking page.

Model	Provider	Strengths	Context
Claude Opus 4	Anthropic	Complex reasoning, coding	200K tokens
Claude Sonnet 4	Anthropic	Balanced performance/cost	200K tokens
Claude Haiku 4	Anthropic	High speed, low cost	200K tokens
GPT-4o	OpenAI	Multimodal, general-purpose	128K tokens
Gemini 2.0 Flash	Google	Multimodal, speed	1M tokens
Llama 3.3	Meta	Open-source, self-hosted	128K tokens

Tuning Strategy Comparison

Prompt Engineering

Best fit: rapid prototyping, small data volumes
Cost: Low
Effect: Moderate

RAG (Retrieval-Augmented Generation)

Best fit: incorporating up-to-date information, injecting domain knowledge
Cost: Moderate
Effect: High (knowledge accuracy)

Fine-Tuning

Best fit: specific style/format, large-scale repetitive tasks
Cost: High (upfront investment)
Effect: High (style consistency)

Fine-tuning is not a single technique. Which one applies depends on what you are trying to move — the task, or the preference:

Technique	What it does	When to reach for it	Relative cost
SFT (Supervised Fine-Tuning)	Trains on labeled prompt/response pairs to teach a task or an output format	You have curated examples of exactly what you want back	Highest — full-weight training
LoRA	Freezes the base weights and trains small low-rank adapter matrices instead	The same goal as SFT at a fraction of the memory; adapters can be swapped per task	Moderate
QLoRA	LoRA on top of a 4-bit quantized base model	Fine-tuning a model on a single GPU that could not otherwise hold it	Low
DPO (Direct Preference Optimization)	Trains on chosen/rejected response pairs to align preference directly, with no separate reward model	Output is already task-correct but the tone, safety, or style is wrong	Moderate

A workable order of escalation: exhaust prompt engineering, then RAG, and only then fine-tune — SFT or LoRA/QLoRA to teach the task, DPO on top to shape preference. QLoRA depends on the quantization choices below, so treat those two decisions as one.

Quantization

Quantization strategies for reducing VRAM usage when self-hosting open-source models:

FP32 → FP16: 50% VRAM reduction, negligible performance loss
FP16 → INT8: another 50% VRAM reduction, minor performance degradation
INT8 → INT4: another 50% VRAM reduction, caution — noticeable performance degradation

Recommended tools: llama.cpp, GPTQ, AWQ, bitsandbytes

Compute Resource Management AI Model Benchmarking