AI Foundations · The 2027 model lineup
You can describe the "free at the base, paid at the frontier" product pattern that every major lab except DeepSeek has now adopted, and name the labs and models on each side of the split.
Refreshed: May 2026.
The 2024 story of open weights was "Meta vs everyone else." Llama was open, the closed labs were closed, and people argued about whether the frontier would migrate one way or the other. By 2026 the story is different and stranger. The pattern is convergence, not conflict — every major lab has converged on the same product structure. Every lab except one.
The pattern: free at the base, paid at the frontier. Each lab publishes a tier of open-weights models intended for fine-tuning, on-prem deployment, edge use, and academic research. The lab's flagship — its hardest-reasoning, most capable model — stays closed. The open tier is a few generations behind the closed one in capability, but it's good enough for most tasks and free to host yourself. The closed tier is the revenue and the safety frontier.
- OpenAI: closed GPT-5.5 / GPT-6 + open gpt-oss (their first open-weights release since GPT-2, shipped 2025). - Google: closed Gemini 3.x / Gemini 4 + open Gemma 4. - Meta: closed Muse Spark + frozen-or-legacy Llama family (Llama 4 / 5 are still downloadable but are no longer the frontier). - Alibaba: closed Qwen 3.6-Plus + open Qwen 3.6 dense and MoE variants.
Chapter contains 4 lessons.