open weight
An AI model whose trained parameters (weights) are publicly released so anyone can download, run, and modify it locally, as opposed to a closed model accessible only via API.
As defined in Ai Sota #2026
Where it comes up · 1
- Ai Sota #202610 claims
Most important claims · 10 of 10Search all claims for “open weight” →
- The number one reason to release an open model is to get people to use it; outside the US many people won't pay for software but have computing resources to run a model locally, and some data shouldn't be sent to the cloud, so open weights enable AI uptake that wouldn't happen otherwise.Ai Sota #2026
- Open models are an engine for AI research because they are what people start with, so it's important to own them; therefore the US should build the best open models so the best research happens in the US and US companies capture the value—and without more investment the ecosystem is dominated by excellent Chinese models (Qwen) cultivating influence, since matching the closed frontier within a half-generation costs only ~$100 million.Ai Sota #2026
- Effectively banning the release of open models over safety risks is impossible without the US building its own Great Firewall (known not to work well), because the $1M-$100M cost of training these models is attainable to a huge number of actors worldwide, so models will be trained everywhere; we want this information and these tools to flow freely.Ai Sota #2026
- Chinese open-weight models are popular partly because their licenses are unrestricted open-source licenses with no strings attached, whereas Llama or Gemma carry restrictions—such as user-count limits requiring financial reporting to the developer—making the Chinese models more freely usable.Ai Sota #2026
- DeepSeek kicked off a movement within China analogous to how ChatGPT triggered a US chatbot wave, producing many strong frontier open-weight labs (Z.ai/GLM, MiniMax, Moonshot's Kimi K2), to the point where DeepSeek is losing its crown as the preeminent Chinese open model maker.Ai Sota #2026
- Whether open models become dominant depends on the trajectory of progress: if saturation comes within a few years while financial support is still strong, open models will be so optimized and cheap to run—because so many people invest in optimizing the serving of common open architectures, even enabling dedicated chips—that they win out over custom closed offerings.Ai Sota #2026
- The big Chinese open-weight models may actually be good for US companies, because seeing that a model like DeepSeek-V3.2 is excellent and used with no backlash or security incident could encourage US labs to release better, more cutting-edge open models than the typically one-generation-behind ones they currently release for safety reasons.Ai Sota #2026
- Chinese companies will keep releasing open-weight models for a few years because there is no clear API business model for them in the US—top US tech and IT companies won't pay for API subscriptions to Chinese firms over security concerns—so open weights become a way to influence and capture a share of the huge growing US AI expenditure market.Ai Sota #2026
- Llama declined because Meta executives, excited by its community popularity, tried to use open source for a bigger splash—building giant Llama 4 models overfit to benchmarks (via preference-trained variants) rather than the small, trustworthy, modifiable models people actually wanted—and it imploded under internal political fighting and misaligned incentives between researchers and management.Ai Sota #2026
- There will be more open model builders throughout 2026 than there were in 2025, with many of the notable ones in China, though building these models is expensive enough that consolidation will eventually occur—just not as a story of 2026.Ai Sota #2026