Chapter 1
Downloadable frontier AI began by accident
Downloadable frontier AI began by accident. In March 2023 the weights of Meta's LLaMA, released only to approved researchers, leaked onto 4chan as a torrent. Four months later Meta made openness a strategy, publishing Llama 2 free for research and commercial use, though under a licence restrictive enough that calling it open source was disputed from the first day. Mistral's Mixtral followed in December 2023 under the genuinely permissive Apache 2.0 licence, and in July 2024 Meta's Llama 3.1 405B became the first downloadable model that independent analysts placed in the same class as the closed frontier.
Chapter 2
Then the centre of gravity moved to China
Then the centre of gravity moved to China. DeepSeek's V3 arrived at the end of 2024, and its R1 reasoning model followed in January 2025 under the MIT licence, priced at about 2% of OpenAI's o1, the model it was pitched against. Within a week the markets had repriced the economics of AI, and a Chinese lab had become, for the first time, the reference point for the frontier. By April 2025 the analyst Nathan Lambert could write that Llama was no longer the open standard, and so it proved: Alibaba's Qwen, DeepSeek, Moonshot's Kimi and Zhipu's GLM became the default foundation of the open ecosystem. OpenAI returned to open weights with gpt-oss in August 2025, its executives saying customers were already running Chinese models. By June 2026 Zhipu's GLM-5.2 was the strongest open model yet released, and the UK AI Security Institute measured the open tier at four to seven months behind the closed frontier on cyber tasks, down from six to ten.
Chapter 3
The summer of 2026 showed how quickly the arrangement could shift
The summer of 2026 showed how quickly the arrangement could shift. In July Meta, whose Llama had made open weights a corporate strategy, shipped its new flagship, Muse Spark 1.1, closed. Days later Moonshot published the weights of Kimi K3, at 2.8 trillion parameters the largest model ever released, and on the Artificial Analysis index, the measure most of this account rests on, the distance between the best downloadable model and the best model overall fell from nine points to four in four days. Meta came back in August with Muse Glimmer, a small Apache-licensed model that runs on a single consumer GPU, and a promise to open its flagship 'soon'. By the end of August the gap on that index was three points, and more than half of the 161 models it scored could be downloaded. Price moved further than capability. DeepSeek's V4-Flash cost about three cents to run the index's full test battery, against $1.86 to $3.15 for the leading closed models, and in late August Alibaba's Qwen3.8-Flash-Next matched GPT-5.5's score at a thirty-third of its input price. SemiAnalysis, measuring across three eras of language models, found that the time open models take to catch up with each new closed advance has halved every generation.
Chapter 4
'Open' has never meant one thing
'Open' has never meant one thing, and the licences show it. Kimi K3's grants almost every right but requires a separate agreement from resellers earning more than $20 million a year. Alibaba requires a paid licence from model-as-a-service businesses above $50 million in revenue to use Qwen3.8-Max, and the weights it published are a text-only, always-reasoning sibling of the model it sells; the score that placed Qwen3.8-Max second among open models was measured on the hosted version. Reach is another limit. The largest models need institutional hardware to run, while Alibaba's Qwen3.8-27B, small enough for a 32-gigabyte laptop, scored level with a model ten times its size.
Chapter 5
In September Artificial Analysis moved to a revised version of its…
In September Artificial Analysis moved to a revised version of its index on which scores run lower, and on it the gap was twelve points: Xiaomi's MiMo-V2.6-Pro led the downloadable tier at 46 against 58 for the best closed model. At the end of October the index was rebuilt again, around private tests and harder coding and science tasks, and open models fell further than closed ones: the best with published weights scored 32, about 27 points behind the leader. The analyst Nathan Lambert judged that the older version had flattered open models; how much of the widening is the models and how much the measure is not settled. Other measures count the gap in months. Mozilla, which advocates open models, put it at 4.4 months, Epoch AI's cyber index at six to seven, and on METR's time-horizon method the best closed model completes tasks 1.7 times as long as the best open one. Labs outside China keep publishing without reaching the top. In October Meta opened the 816-billion-parameter Muse Spark Open under its own restrictive licence, the best American open model; Reflection AI kept its promise to publish Beam's weights under Apache 2.0; and Mistral published its trillion-parameter Large 4. Whether the lag settles where Lambert forecast in February, six to nine months behind by the end of 2026, is the open question.
And now
The gap between the best downloadable model and the best closed one more than doubled on paper this week, because the ruler changed. On 27 October Artificial Analysis rebuilt its Intelligence Index around private and harder tests. The best models with published weights, DeepSeek's V4.1 Flash 1017 and Zhipu's GLM-5.3, score 32 on it, 26 points behind the leader; a day later Google's closed Gemini 4 Carbon took the top at 59. On the old version the gap was twelve. The analyst Nathan Lambert says the old index flattered open models; Epoch AI's cyber measure, a different method, puts the best open model six to seven months behind.
Two labs published weights they had promised. MiniMax released M3.1 Flash on 23 October, four days after launch; it scores 29. Mistral released Large 4, about a trillion parameters, on 29 October, inside its end-of-month promise, under a licence that charges hosting businesses; it scores 26.
Moonshot's Kimi K3.1 Pro scored fourth of all models at its 23 October launch, and the index said its weights were expected soon after. They have not appeared. Meta's Muse Spark Open, at 25, is still the best American open model.
Where it stands today →