Two American AI labs released open-weight models this week in a bid to claw back ground from Chinese developers who now dominate the field — and the stakes are nothing less than who builds the backbone of the next century's surveillance and intelligence infrastructure.
Reflection AI dropped Beam, a 501-billion-parameter model, on the same day Meta rolled out Muse Spark Open, an 816-billion-parameter model it calls "the most capable US open-weight model to date." Both are aimed at the open-weight market where Chinese labs — Alibaba's Qwen, Moonshot AI's Kimi K3, Z.ai's GLM, and DeepSeek — have been eating American lunch for two years.
Here's the problem: by each company's own published scores, neither American model matches the top Chinese models on Terminal-Bench 2.1, a widely used test of agent tasks. The U.S. is playing catch-up on its own soil.
Reflection's pitch is efficiency, not brute force. The company says Beam completes tasks in far fewer tokens than competitors, which drives down the real cost per finished job even if per-token pricing stays the same. Reflection co-founder and CTO Ioannis Antonoglou told VentureBeat the model was trained with a penalty for unnecessary tokens: "You don't give up anything. You just penalize it in terms of how many tokens it can use." The trick, he said, is applying that penalty without the model simply quitting — which requires massive reinforcement learning. Reflection says it ran over 100 million trial-and-error rollouts across a fleet of 10,500 Nvidia GB300 GPUs and peaked at 883,006 virtual CPUs on Google Cloud and AWS. Independent benchmarking firm Artificial Analysis says early indicators suggest Beam will be "one of the most token-efficient open models we've seen for its level of intelligence."
But DeepSeek — the Chinese lab most praised for efficiency — is conspicuously absent from Reflection's own comparison chart. Draw your own conclusions.
Meta, meanwhile, is selling capability. The Guardian reported that Muse Spark Open is "built for coding and agentic work" and is optimized to run on Apple's M5 Ultra chips — an unusual partnership between the two tech giants. Meta says users can run the model on two Mac Studios to create a desktop "supercomputer." Meta also says "strong safeguards" have been integrated into the new model. Whenever a tech company volunteers that phrase, ask what speech gets walled off and who decides.
The open-weight distinction matters. Unlike closed models like ChatGPT or Claude, developers can download open-weight parameters and modify them. That's the kind of decentralized, user-controlled architecture that keeps power out of institutional hands — which is exactly why who dominates the space matters. VentureBeat noted that many U.S. companies are already embracing Chinese open models because they're cheaper. Every American enterprise that builds on Beijing's infrastructure is a node in a dependency chain controlled by a foreign rival.
The question isn't whether Beam or Muse Spark wins a benchmark today. It's whether American labs can build models good enough and open enough that U.S. enterprises stop handing their architecture — and their data — to Chinese competitors. Right now, the scoreboard says they're not there yet.
