Every open model, measured against the frontier

52 open-weight models you can download and run at home, ranked on the Artificial Analysis Intelligence Index v4.3, with the hosted models from Anthropic and OpenAI dropped into the same table for scale. Each model has one row, at the lightest build this site prices, and two of them are also priced at a heavier quantisation, named in the row itself. Each row links to what it takes to run it, and each price opens the calculator on that machine running that model. For which machine pays back soonest at each level, see best buys by usage; for two of them side by side, every head-to-head; for all of them at once, every machine on the site in one table.

Jump to: How far behind the frontier open models are · Every model here, ranked · Models the index has not scored yet · How to read the scores

How far behind the frontier open models are

The short version: the best open model here scores 43 — that is MiniMax M3.1 Flash, and it wants 249 GB of memory. The best hosted model scores 53. That gap of 10 points is the thing no amount of hardware closes.

Every model here, ranked

Each machine named is the cheapest that holds that model at the 32k context the calculator starts at, of the 37 machines here still sold at a published price. Tencent Hy3 fits nowhere at that length: its row names the machine that holds it at 16k, which is marked beside the price. MiniMax M3.1 Flash fits nothing still sold at any window: the two machines here that hold it are one no longer sold and one with no published price, and its own page prices them. Ling 3.1 flash fits nothing still sold at any window: the two machines here that hold it are one no longer sold and one with no published price, and its own page prices them. Beam 1.0 501B-A23B fits nothing still sold at any window: the two machines here that hold it are one no longer sold and one with no published price, and its own page prices them.

Models the index has not scored yet

17 more open models on this site are not in the table above: Spark-X2.6 1.7B, LFM3 1.7B, Spark-X2.5 4B, Spark-X2.6 4B, Ornith 1.5 9B, LFM2.5 24B-A2B, Arete-27B, Hy-Memory 30B-A3B, Laguna XS 2.1, Nanbeige4.3 33B, KAT-Coder V2.5 Dev 35B-A3B, Ornith 1.5 35B-A3B, Clover 35B-A3B Preview, Flint 40B-A2.5B, Ripple 65B Preview, Ornith 1.5 122B, GLM-5.3-SLIM. Their pages show what each one needs and what runs it.

How to read the scores

Marked 'estimated' where Artificial Analysis imputed the score rather than running the full v4.3 suite. 26 of the models here carry that mark. Scores are the Artificial Analysis Intelligence Index v4.3, read between 2026-09-11 and 2026-10-27. Artificial Analysis Intelligence Index v5.0.1 has replaced v4.3 at the source, on a different set of evaluations, so a score on it is not comparable with one here. Read on 2026-11-08 it covered 10 of the 54 models ranked on this site and not the other 44, so the figures here stay on v4.3, the one scale that covers all of them. Hybrid models are shown at their reasoning or highest-effort score, with the alternative noted on each model's page and, on the hosted rows above, beside the score itself. Weights are the download; a running model also needs a cache the size of your context window, so see how much memory each size really takes. 25 of these models read images as well as text, and there the download is the language model alone: 16 of their pages price the vision projector that loads beside it.