Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost
Reflection AI has announced Beam, an open-weight, text-only mixture-of-experts model the Brooklyn startup says matches top Chinese open models on advanced reasoning tests while using substantially less compute. Beam is a 501-billion-parameter model with 23 billion active parameters, trained on 23.8 trillion tokens and offering a 1 million token context window; Reflection plans to publish weights and technical details this month.

Why It Matters
If Reflection's performance and efficiency claims hold up, Beam could accelerate adoption of high-end open models outside closed offerings from companies like OpenAI and Anthropic, and support enterprises and governments building customized local AI systems using proprietary data.
Key Facts
- Model name: Beam
- Company: Reflection AI (Brooklyn-based startup)
- Parameters (total): 501 billion
- Active parameters: 23 billion
- Pretraining tokens: 23.8 trillion tokens (pretraining)
Reflection AI unveiled Beam, its first frontier open-weight model, positioning it as a lower-compute alternative to leading Chinese and Western open models. Reflection described Beam as a text-only mixture-of-experts model optimized via high-compute reinforcement learning to improve reasoning, coding, and agentic tasks while reducing token cost and inference-time compute. Beam is reported to contain 501 billion parameters with 23 billion active at inference, and it was pretrained on 23.8 trillion tokens. The model supports a 1 million token context window. Reflection contrasted Beam with Z.ai’s GLM-5.2 — which it said has roughly 744 billion total parameters and 40 billion active — and claimed Beam matches GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. The startup said Beam outperforms current leading Western open models on the same reasoning tasks, and provided benchmark comparisons versus other recent releases: Reflection’s results show Beam scoring higher than Inkling on four coding tests where both reported outcomes, though Inkling is multimodal and Beam is strictly text-only. Reflection’s performance claims have not been independently verified. Reflection is marketing Beam and future models to enterprises and sovereign customers through an "AI factory" concept that would let institutions fine-tune Reflection’s models on their own proprietary data to produce localized systems. The company has already pursued commercial and sovereign partnerships, including testing a sovereign AI factory with South Korea’s Shinsegae Group, and says it will distribute Beam’s weights and technical documentation this month through hyperscalers, neoclouds, and open source integrations. Founded in 2024 by two former Google DeepMind researchers, Reflection has raised about $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, and was last valued at a $25 billion pre-money figure according to PitchBook. The startup has also secured large compute arrangements this year, signing deals with SpaceX and Nebius worth more than $7 billion combined to gain access to Nvidia GB300 chips through 2029.
Keep Reading

Instinct brings its AI agent to group chats, even for friends without an account

After Factory’s public spat with Khosla, Menlo proudly invests
This new AI model could help America close a technological gap with China
