There is no “AI.” There are models.

When someone says “AI is dangerous” or “AI will replace radiologists” or “AI uses too much water,” the sentence smuggles a unified subject that does not exist. A 39M-parameter sentence-embedding model is AI. A 2GB Whisper transcription model is AI. A 671B MoE reasoning model is AI. They share almost no engineering, no training cost, no failure mode, and almost no user.

The singular word does work for the people selling the strongest closed tier — it lets the public conversation inherit the prestige of every prior generation while the actual business model is monetizing one specific row. It does not do work for anyone trying to think about the field clearly.

Models are not truths. They are weights produced by training a specific architecture on a specific dataset toward a specific objective. Every part of that sentence is a choice, every choice is contestable, and every choice changes when the next dataset lands. The map below is a snapshot — explicitly so.

The honest first move when reading any AI claim is to ask which row it is about. “Models hallucinate” applies to row 1. “Models are deterministic” applies to rows 3 and 4. “Models are stochastic parrots” is a critique of one specific architecture, generalized rhetorically to all of them. Each row has its own answer.

Two structural reasons the gap is narrowing. First, training cost dropped: DeepSeek V3 and successors made it visible that frontier-class capability costs an order of magnitude less than the public US-lab figures assumed. Second, the Chinese open-weights labs — DeepSeek, Moonshot, Z.ai, Alibaba, MiniMax — have made a strategic choice to release weights, not just papers. That puts the same models in the hands of a researcher in Lagos and a hyperscaler in Seattle.

One reading of this chart: the conversation that frames AI as a contest between three or four firms in California is already two years out of date.

The trend across 2026 has been toward Apache 2.0 as the default for new releases — Mistral, Qwen, GLM, MiniMax, FLUX Klein. Google moved Gemma 4 to Apache 2.0 in March, the first time a US frontier lab’s open tier went fully OSI-approved. The exceptions — Llama’s community license, FLUX.1 dev’s non-commercial clause, Stability AI’s under-$1M-revenue gating — are now the read-the-fine-print cases rather than the norm.

A note on responsibility licenses (OpenRAIL family). StarCoder 2’s license forbids certain downstream uses — biometric ID, surveillance, generation of disinformation. These restrictions are largely unenforceable in practice; their function is closer to a public-norm signal than a contractual one. Reasonable people disagree on whether that’s the right move; it is at least an explicit one.

Some structural notes on the benchmarks themselves. MMLU is saturated — every frontier model scores 88–92, the differences are within run-to-run noise. GPQA Diamond is the current general-knowledge ceiling test. SWE-Bench Verified is the most defensible code benchmark (real GitHub issues, repo-level edits, no leakage from the pretraining set). Chatbot Arena Elo measures preference, not correctness — a polite hedger can beat an accurate blunt model. GDPval-AA is the newest entrant: real knowledge-work tasks across 44 occupations.

A benchmark that no model has saturated is informative. A benchmark every model has saturated is informative only about the benchmark.

The grid above is incomplete by construction. Hundreds of fine-tunes, dataset curators, eval-harness maintainers, and vendor-neutral host organizations sit downstream of every model listed here. The labour is dispersed; the singular noun “AI” is the only thing that compresses it back into a single subject.

Tools evolve faster than systems do. The systems keep optimizing as though the tools hadn’t changed, until they’re stuck in scale problems we couldn’t foresee. When every inherited truth is shaking at once, we believe the responsible response is neither to manufacture new certainty, nor to refine the systems blindly. It is to return to the why — to the original problem they were meant to answer, before the solution itself became the silo.

Calling the entire field “AI” is one such silo. It is a cognitive convenience that hides which architecture is being discussed, which dataset is being scrutinised, which community is being either credited or erased. The map on this page does not claim to dissolve the convenience. It tries to make it visible — so the conversation that follows can be about an actual row.

Models are not truths. They are weights, trained on data, designed to change as data changes — unsteady by construction. The work this page tries to honour is the work of the people training, evaluating, releasing, and maintaining them in the open. Most of that work is not done by the firms whose names appear in the headlines.