LLM Model LandscapeIntermediate

Llama Model Size/Performance Chart

The Llama lineup and licence, serving with Ollama and vLLM, GPU memory sizing, and an honest look at when a hosted API is still the better choice.

By Amit UpadhyayPublished 13 Oct 202611 min

Meta's Llama is the most popular open-weights model family, enabling self-hosting with no per-token fees. Llama 4 (Maverick and Scout, April 2025) is the newest open-weight generation, and Llama 3.3 70B-Instruct (Dec 2024) remains the most widely deployed. Meta's newer Muse Spark models are closed-weight, so the open line has not moved since Llama 4 — check it against newer open models (DeepSeek V4, Qwen 3.8, Mistral Large 3) before committing. Models are available on HuggingFace for download. Llama uses a custom commercial license (Meta Llama Community License) allowing commercial use for apps with <700M monthly active users. Key use cases: privacy-critical apps, cost-sensitive high-volume, fine-tuning on proprietary data, air-gapped deployments.

The pipeline

Each stage in order — click any step to read what it does.

Interactive Pipeline Flow

UnderstandImplementOptimise

📥 Download — HuggingFace hub

Download from HuggingFace: meta-llama/Llama-3.3-70B-Instruct (requires Meta approval). Or use Ollama: ollama pull llama3.3. Llama 3.2 1B/3B/11B available without approval gate.

Architecture diagrams

Llama family by size and deployment target.

Loading diagram...

Key design decisions

The trade-offs worth knowing before you build this.

🔒Llama enables private, air-gapped AI deployments

For healthcare, legal, and government use cases where data cannot leave the network, Llama + vLLM on private infrastructure is the only viable option. No API keys, no data egress, no vendor dependency.

🎯A fine-tuned small Llama can beat a general model on narrow domains

A Llama 3.1 8B fine-tuned on 2000 domain-specific examples can outperform a much larger general-purpose model on that specific task. The cost argument is narrower than it was now that hosted economy models cost $0.10 per million input tokens, so run the numbers. For repetitive, well-defined tasks, invest in fine-tuning.

⚡Specialised hardware makes Llama inference very fast

Groq LPU hardware runs Llama 3.3 70B at 350+ tokens/second, several times the speed of typical hosted frontier models. For latency-critical apps (real-time voice, streaming chat), Groq + Llama delivers the best perceived speed at a fraction of the cost.

Next: Animate →

Llama open-weight self-hosting vLLM quantization

← PreviousEvals for Choosing a Model Next →Claude: Anthropic's Model Family in Practice

Continue reading

All LLM Model Landscape articles →

LLM Model Landscape11 min

GPT: OpenAI's Model Family in Practice

The GPT-6 tiers, what reasoning effort replaced, when to use the Responses API over Chat Completions, how the bill is calculated, and what is different on Azure.

LLM Model Landscape11 min

Claude: Anthropic's Model Family in Practice

The four Claude tiers, adaptive thinking and effort, the tool-use rules that differ from other providers, prompt caching, and where you can run it.

LLM Model Landscape10 min

Evals for Choosing a Model

Why public benchmarks do not transfer to your workload, how to build a model-selection eval, and why cost per successful task beats price per token.

Discussion