Reflection AI launches Beam open-weight reasoning model
Reflection AI has unveiled Beam, its first frontier open-weight AI model: a text-only mixture-of-experts system with 501 billion total parameters, 23 billion active parameters and a one-million-token context window. The two-year-old startup says Beam matches leading Chinese open models on advanced reasoning benchmarks while using three to four times less inference compute.
Beam was pre-trained on 23.8 trillion tokens and trained with high-compute reinforcement learning. Reflection positions the model for reasoning, coding and agentic tasks, describing it as a workhorse for enterprises, the public sector and developers. The company plans to release the model weights and full technical details this month.
A challenge to Chinese and Western open models
Reflection says Beam performs on par with Z.ai's GLM-5.2 in advanced reasoning evaluations and outperforms current leading Western open models. GLM-5.2 has roughly 744 billion total parameters and 40 billion active parameters, giving Beam a smaller active footprint by that comparison. The benchmark results and efficiency claims have not been independently verified.
The company is positioning Beam against closed-model providers Anthropic and OpenAI, as well as Chinese developers including DeepSeek, Qwen and Z.ai. It also names Western open-model competitors such as Mistral, Meta and Cohere. Reflection's own reported results show Beam ahead of Thinking Machines Lab's Inkling on four coding tests reported by both companies, but the comparison has limits: Inkling is multimodal while Beam is text-only.
Compute commitments underpin the launch
Training and serving frontier models depends on sustained access to specialised infrastructure. In that context, Reflection's Nebius compute strategy illustrates Reflection's compute strategy alongside its broader effort to secure Nvidia GB300 capacity through 2029.
Reflection said this summer that its agreements with SpaceX and Nebius were collectively worth more than $7 billion. The startup, founded in 2024 by two former Google DeepMind researchers, has raised roughly $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners, according to PitchBook. Its latest funding round valued the company at $25 billion pre-money.
Local AI systems are the commercial target
Beyond the model release, Reflection is promoting “AI factories” that would allow institutions to train customised local systems using their proprietary data. The company is aiming this proposition at enterprises and sovereign nations and has begun testing a sovereign AI factory partnership with South Korea's Shinsegae Group. Axios reported that hedge funds and trading firms are among organisations interested in such systems.
Beam is expected to be distributed through hyperscalers and neoclouds, with integrations across open-source libraries available at launch. For buyers, the business implication is to assess released weights, practical inference costs, latency and deployment controls against their own workloads before relying on vendor-reported benchmark comparisons.