US and European open-weight models trail Chinese rivals like GLM-5.3

The announcements from Reflection and Mistral come as the US and Europe try to regain ground in open-weight AI. The US has largely fallen off the open-weight frontier, partly because Meta stopped making good LLaMA models. Reflection's Beam, despite being trained on 6,144 GB300 NVL72 GPUs for four weeks, still trails Chinese models like GLM-5.3. Mistral, described as Europe's leading AI developer, plans to release Mistral Large 4 by the end of the month. Its scores show it will likely be the strongest Western open-weight model upon release, but still less powerful than Chinese ones.

The competitive gap is evident in benchmark results. Beam achieves token efficiency—comparable to GLM-5.2 while using 3–4× less inference compute—but its raw scores on DeepSWE 1.1, GPQA Diamond, and MCP Atlas are lower than GLM-5.3. Mistral Large 4 is preferred in CAD and STEM tasks in internal tests, but performs on par or close to GLM-5.3 in finance and coding. These results suggest that while Western labs are making strides, catching up to Chinese open-weight models remains a challenge.