A $6.3 billion open-weight model just got embarrassed by the French...

Summary

The broadcast opens by examining a wave of global competition in the open-weight artificial intelligence ecosystem. It highlights how growing interest in sovereign artificial intelligence stems from privacy and security risks associated with proprietary models. In recent examples, user conversations logged by proprietary providers like Claude prompted safety escalations and police interventions. For enterprise and governmental infrastructure such as electrical power grids, sending proprietary schematics and telemetry to third-party providers presents intolerable risks of data interception, data breaches, and industrial espionage, creating intense demand for models that can run locally on private hardware.

In response to this demand, Reflection AI, an Nvidia-backed startup valued at 25 billion dollars that raised 4.6 billion dollars in funding, announced its first major open-weight release named Beam. The model contains 501 billion total parameters and 238 billion active parameters. Reflection AI trained Beam's base weights in under four weeks on an in-house cluster of 6,144 Nvidia GB300 NVL72 GPUs across approximately twenty-four trillion tokens, leveraging a compute agreement valued at up to 6.3 billion dollars. The team then spent an additional four weeks post-training the model to specialize in coding benchmarks, tool use, and long-horizon autonomous agent tasks, matching larger models while consuming substantially less inference compute.

Twenty-four hours later, French artificial intelligence laboratory Mistral released Mistral Large 4, a mixture-of-experts model containing one trillion total parameters with forty-nine billion active parameters per token. Mistral Large 4 features native multimodality, a one-million-token context window, and inference efficiency designed to emulate a much smaller model. On the DeepSWE 1.1 benchmark, Mistral Large 4 scored 62, outperforming Beam's score of 44 by eighteen points and surpassing GLM-5.3. It also placed in the top five on the Artificial Analysis Cyber Index for identifying cybersecurity vulnerabilities, an application Mistral specifically targets for enterprise utility and government defense contracts.

These Western developments unfolded as Chinese artificial intelligence provider Moonshot AI completed a funding round at a 50-billion-dollar valuation while preparing an initial public offering in Hong Kong. Moonshot's flagship 2.8-trillion-parameter open model, Kimi K3, had previously seen such intense demand that new user subscriptions were paused due to hardware capacity limits. For practitioners selecting among these options, the video recommends Google's Gemma 4 for single-machine consumer hardware, Kimi K3 for high-capacity unconstrained data centers, and Mistral Large 4 for enterprise environments subject to compliance oversight. Finally, the segment highlights Kernel's Config Registry, an open-source browser infrastructure tool that enables autonomous agents to bypass web bot checks.