Pre-trained AI Models
Pre-trained model families split open-weight vs. closed-API. Meta's Llama 4 Scout (17B active/109B total MoE) gives a 10M-token context window under a custom Community License, free for commercial use below 700M monthly active users; Maverick (400B total) caps at 1M tokens. Google's Gemma 4 moved from Gemma 3's restrictive Terms of Use to full Apache 2.0, spans E2B to 31B parameters with up to 256K context. Mistral Large 3 (41B active/675B total MoE, Apache 2.0) and Small 4 hit 256K context and are freely self-hostable and fine-tunable. Closed models trade weights for scale: Anthropic's Claude Opus 4.8 ($5/$25 per million tokens) and OpenAI's GPT-5.5 ($5/$30 per million tokens) both ship 1M-token context via API only, no downloadable weights, no local fine-tuning. Pricing: Open-weight families (Llama 4, Gemma 4, Mistral Large 3/Small 4) are free to download and self-host under Apache 2.0 or Meta's Community License; hosted frontier APIs run $1-$30 per million tokens (GPT-5.6 Luna $1/$6, Claude Opus 4.8 $5/$25, GPT-5.5 $5/$30 input/output per MTok).
Pick open-weight (Llama 4, Gemma 4, Mistral) when you need to fine-tune on proprietary data, run air-gapped, or dodge per-token bills at scale — Mistral's Apache 2.0 license has zero usage caps, unlike Llama 4's 700M-MAU trapdoor that forces enterprise negotiation. Pick closed APIs (Claude Opus 4.8, GPT-5.5) when you need frontier reasoning today and can't afford your own GPU cluster — self-hosting a 675B-parameter MoE model is not a weekend project. Even Mistral admits Small 4's configurable reasoning effort trades latency for accuracy; there is no free lunch, whatever the benchmark charts claim. Known weakness: You can't get both: the highest-capability models are API-only with zero weight access, while the openly licensed models require you to run your own GPU infrastructure to realize their advertised context windows.