GPT-4 / GPT-4o
OpenAI, closed
Claude
Anthropic, closed
Gemini
Google, closed
Llama
Meta, open weights
Mistral/Mixtral
Mistral, open weights
DeepSeek
DeepSeek, open weights
Qwen
Alibaba, open weights
Gemma
Google, open weights
The specific names on top of any 'best model' leaderboard change every few months, so memorizing today's winner is a losing game. What's stable is the set of *axes* along which these models actually differ, and understanding those axes lets you evaluate whatever model comes out next without starting from scratch.
- Open weights vs. closed/API-only — can you download and run the model yourself, or only call it through a paid API?
- Dense vs. Mixture-of-Experts (MoE) — does every token pass through the whole network, or only a routed subset of specialized sub-networks?
- Parameter count and training compute — rough proxies for capability, though far from perfect ones today.
- Context window — how many tokens of input the model can attend to at once.
- Modality — text-only, or natively multimodal (text, image, audio, video)?
Closed, API-only models (GPT-4-class, Claude, Gemini) are typically the largest and most capable at launch, updated frequently, and require no infrastructure to use — but you're dependent on a vendor's pricing, uptime, and policies, and can't inspect or modify the weights. Open-weight models trade some peak capability for control: you can fine-tune them, run them privately, and inspect (though not fully understand) what's inside.
- Bigger dense models cost more to serve per token than smaller or MoE models of similar capability.
- Context window claims ("1M tokens!") don't always translate to actually using that context well — see the evaluation star for why raw specs mislead.
- Multimodal-native models are trained from the start on mixed text/image/audio data, rather than bolting a vision encoder onto a text-only model after the fact.