What TrustList's AI-model board actually knows about 88 models, and what it doesn't yet

TrustList's AI-model board currently tracks 88 models. Look closely, though, and it is really two boards in one: 24 models carry a full technical specification - provider, context window, licence, pricing where public - and 64 do not yet. This piece is about the 24, and about being honest that the other 64 exist.

Who's represented, among the ones we can actually describe

Restricting to the 24 fully-specified models, the provider spread is:

Provider	Models with full specs
Alibaba (Qwen)	6
Z.ai (Zhipu AI)	4
IBM	4
Anthropic	2
DeepSeek	2
xAI	2
Google	2
Moonshot AI	2

That is not a claim about the whole 88-model board's ownership - it is a claim about which 24 happen to be fully documented right now, and those 24 skew toward the models added or corrected most recently. Treat it as a snapshot of documentation coverage, not market share.

Two real patterns inside those 24

Context windows have converged. Of the 24 models with a known context window, 18 - three out of four - carry 1,000,000 tokens or more. A 1M window was a headline eighteen months ago. Among the models TrustList has fully documented today, it is closer to the default.

Open weights are a real minority, not a fringe. 10 of the 24 either publish weights directly or carry a licence naming Apache, MIT or an explicitly open designation. That is a genuine option, not a rounding error

• roughly two in five of the models with a full spec can be run on infrastructure the buyer controls rather than called through an API.

What "18 have a benchmark score" actually means

Only 18 of the 88 models on the board carry a benchmark score of any kind. Every one of those scores is stored with is_verified: false and a source_url pointing at where the number came from - a policy decision made explicitly because a directory that presents scores as fact without saying where they came from is not trustworthy, whatever the scores say. A missing score on this board is not evidence a model is weak. It is evidence nobody has entered a sourced number for it yet.

Reading the board honestly

Two-thirds of the models here don't yet have a documented spec, and four-fifths don't yet carry a benchmark. That is not a flaw to route around quietly - it is the actual, current state of the catalogue, and a buyer comparing models on this board should read a blank spec field as "not yet documented," never as "nothing to report." The 24 fully-specified models are a reliable sample of what a modern frontier or open-weight spec sheet looks like. They are not yet a complete picture of the other 64.

---

Figures pulled from TrustList's live AI-model board (/benchmarks/ai-model-board) on 31 August 2026: 88 models total, 24 with a populated model_spec, 18 carrying at least one benchmark score.