Reflection AI launches open-weight reasoning model Beam with NVIDIA backing

Reflection AI has shipped Beam, a 501-billion-parameter open-weight model focused on complex reasoning and coding. NVIDIA highlighted the launch and the companies say Beam reaches frontier-level reasoning benchmarks at substantially lower inference cost on NVIDIA GPU infrastructure. Some outlets have called it the 'DeepSeek of the West'.

The framing is deliberate. Open-weight leadership has shifted toward Chinese labs, and DeepSeek's V4.1 Flash has narrowed the overall US-China gap to about 3% by Bloomberg Intelligence's measure. Beam is an attempt to give US enterprises and governments an open model they can self-host without Chinese-origin concerns. NVIDIA's involvement fits its strategy of encouraging strong open models that drive demand for its hardware and software stack.

Beam lands on the same day as Mistral Large 4, making October 6 a crowded day for open weights. Both arrive just as Meta signals it will be more selective about open-sourcing after Llama 4's problems, which leaves room for new non-Chinese open leaders.

The Hacker News reception (541 points) was interested but cautious, with developers weighing whether the benchmarks hold up independently and what the license allows. Watch for third-party evals against DeepSeek V4.1 Flash and Mistral Large 4, availability on major clouds, and real inference-cost numbers once people deploy it.