The Modern Model Zoo

The LLM landscape splits roughly into three camps. Knowing which model is which — and what it's optimised for — saves you money and shipping headaches.

The closed frontier (proprietary)

These labs control their weights and serve via API only:

• OpenAI — GPT-4.1, GPT-4o, o3-mini, o3. Generalist powerhouse, frontier reasoning, multimodal (text + image + audio + video). Premium price.
• Anthropic — Claude 4 Opus, Claude 4 Sonnet, Claude 4.5 Haiku. Strongest on coding and long-context reasoning. Constitutional AI alignment.
• Google DeepMind — Gemini 2.5 Pro, Gemini 2.5 Flash. Largest native context (1M+ tokens), excellent multimodal, generous free tier via AI Studio.
• xAI — Grok 3, Grok 4. Real-time X data, looser content policies.

The open-weight champions

Weights are downloadable. You can self-host, fine-tune, or just use the cheap API:

• Meta Llama — Llama 4 Scout (16-expert MoE, 512K context), Llama 4 Maverick (128-expert MoE, 1M context). The default open foundation model.
• DeepSeek — DeepSeek V3, DeepSeek R1 (reasoning). Chinese lab, frontier-quality open weights, extraordinary value.
• Mistral — Mistral Large 2, Mistral Small 3.1, Codestral. Strong European alternative, strong code, multilingual.
• Alibaba Qwen — Qwen3 235B, Qwen3 32B, QwQ. Top multilingual + reasoning open models.
• Google Gemma — Gemma 3 27B, 12B, 4B. Open-weight Gemini cousins.
• Microsoft Phi — Phi-4. Tiny models trained on heavily curated synthetic data.

The reasoning specialists

A new category that emerged in 2025: models trained with extensive RL to think before answering, producing visible chain-of-thought:

• OpenAI o3, DeepSeek R1, Qwen QwQ, Google Gemini Thinking

These are slower and more expensive per query, but blow past standard models on math, code, and multi-step problems.

How to choose

A rough heuristic:

You'll go deeper on model selection in Level 4. For now: there's no single "best" model. Match the model to the job.