Coding Horizon

The Best Open AI Models Are The Ones Nobody Trusts

Every figure, name, date and price the finished picture puts on screen, chased to a source. Anything that could not be chased is listed at the bottom and is not rendered.

The warning the video opens on

Arena CEO Anastasios Angelopoulos says enterprises do not know which models to trust, and are as wary of Chinese open weight models as they are of the frontier labs.

His words: “It’s not only true that they’re terrified of working with the frontier labs, but they’re also terrified of working with the Chinese open source.” He also said enterprises are “in a really tricky situation … just the way that the cards have landed”.

Arena is the evaluation company behind the public model arena, which is why its CEO sees which models enterprises are actually shortlisting.

The Chinese open weight labs the video names

Moonshot AI, DeepSeek, Alibaba and Z.ai. All four publish downloadable weights.

Kimi K3 (Moonshot AI)

2.8 trillion total parameters as a mixture of experts, 16 of 896 experts active per token, a 1,000,000 token context window, open weights released 27 July 2026. At release it was the largest open weight model published.

DeepSeek R1

Released 20 January 2025 and the moment open weight reasoning became a mainstream enterprise question rather than a research one.

Qwen (Alibaba)

Passed 700 million cumulative downloads on Hugging Face, announced by Alibaba on 13 January 2026, overtaking Llama as the most downloaded open model family. Derivative models built on Qwen checkpoints passed 200,000 later the same month.

GLM 5.2 (Z.ai)

Open weights under the MIT licence, 1,000,000 token context window, built for long horizon coding and agent work, released June 2026.

The price gap the cost chapter draws

GLM 5.2 is listed at $4.40 per million output tokens. GPT 5.5 is listed at $30.00 per million output tokens. That is roughly a seven times difference on the same unit, which is the shape of the price pressure the video describes.

The policy fight

Whether Chinese open weight models should be restricted in the US is unsettled and is being argued in public by the industry and by Washington at the same time.

On 24 July 2026, twenty five companies including Nvidia, Microsoft, Meta, IBM and Palantir published an open letter, “Open Weights and American AI Leadership”, urging policymakers not to impose broad restrictions on open weight models. Anthropic has argued the other way, for tighter restrictions on Chinese AI. China’s own commerce ministry has separately been drafting export controls on model weights.

Why downloaded weights do not phone home

Weights are data, not a program. A downloaded checkpoint has no network client in it and no way to initiate a connection; whatever talks to the network is the runtime you chose and the dependencies you installed around it. That is why the video puts the risk in the deployment rather than in the file.

The qualification that matters, and the reason the video keeps insisting the stack is the real surface: loading a model can execute code. Python pickle based checkpoints run arbitrary code at load time, and malicious models exploiting exactly that have been found on public model hubs. The safetensors format exists to remove that path by storing only tensors.

Running weights locally

The pull the video shows on screen is the real command form for a local runtime.

Not chased to a primary