Open-Weight AI Models: This Week’s Race

Why open-weight models matter

• With proprietary AI, prompts go to a company’s servers, where they may be filtered, reviewed, or used according to the company’s policies.
• Open-weight models can be run locally or on a country’s own infrastructure, helping protect sensitive personal and government data.
• That control is driving governments around the world to build or adopt models they can trust.

The models in focus

Reflection AI’s Beam

• A 501-billion-parameter model trained on roughly 24 trillion tokens, followed by additional post-training for coding, tool use, and long-running agent tasks.
• Reflection says it performs comparably to GLM 5.2 while requiring a quarter of the hardware.
• At the time of the video, Beam required an invite, and its weights had not yet been released. The video says they were expected later that month.

Mistral Large 4

• A mixture-of-experts model with about one trillion total parameters, of which 49 billion are active for each token.
• Mistral claims a million-token context window and strong performance on coding and security tasks.
• The video reports that it beats Beam on its cited benchmarks, but notes that the model was still in training, the results came from Mistral, and the weights were due later that month under a custom license.

Moonshot AI’s Kimi K3

• A 2.88-trillion-parameter open model released in July.
• Its popularity reportedly caused Moonshot to stop accepting subscribers after running short of GPU capacity. Microsoft, Amazon, and Google were said to be negotiating to host it.
• Moonshot closed a funding round at a $50 billion valuation and was preparing for a Hong Kong IPO.
• The video also mentions Anthropic’s allegation that the company used fake accounts to make 300,000 requests to copy Claude.

Gemma 4

• Google’s model, released under Apache 2.0 in April.
• The video identifies it as the strongest American open model actually available at the time.

Which model should you use?

• Gaming PC: Gemma 4.
• Data center, without a compliance department: Kimi K3.
• Data center with compliance requirements: Mistral Large 4.
• Want an American option: Watch for Beam’s weights to be released.

Sponsor: browser infrastructure

The video also promotes browser infrastructure for AI agents, including a configuration registry that recommends browser and proxy settings for accessing specific sites. It says the service tests setups against live websites and provides success rates, sample sizes, and verification dates.