started · updated
Reflection AI launches Beam model to challenge Chinese AI dominance
Reflection AI, a Brooklyn-based startup founded by former Google DeepMind researchers, has launched Beam, an open-weight artificial intelligence model designed to compete with leading Chinese models like DeepSeek and Alibaba's Qwen.
Beam utilizes a sparse Mixture-of-Experts (MoE) architecture with 501 billion total parameters, of which only 23 billion are active per request. This design aims to provide high-level reasoning and coding capabilities while significantly reducing inference compute costs—estimated at three to four times less than some competitors. The model was trained on 23.8 trillion tokens and features a context window of up to 1 million tokens.
Supported by major investors including Nvidia and Sequoia Capital, Reflection AI has reached a valuation of approximately $25 billion. To support its massive training requirements, the company has secured significant compute resources through deals with SpaceX and Nebius. The company plans to release the model weights under an Apache 2.0 license later this month, positioning Beam as a cost-effective, high-performance alternative for enterprises and developers seeking to build localized AI systems.
Entities
Alibaba · Beam · DeepSeek · Misha Laskin · Nvidia · Qwen · Reflection AI
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 11 SOURCES] Reflection AI launched Beam, a 501-billion-parameter sparse Mixture-of-Experts model. almakos.com · cybernoz.com · www.prodiris.fr · www.sbctv.gr · futurumgroup.com · +6 more
- [● 4 SOURCES] The model weights will be released under an Apache 2.0 license. cybernoz.com · futurumgroup.com · www.generation-nt.com · news.cnyes.com
- [● 4 SOURCES] Beam performs similarly to Z.ai's GLM-5.2 on advanced reasoning benchmarks. almakos.com · www.sbctv.gr · au.pcmag.com · news.cnyes.com
- [● 3 SOURCES] Reflection AI has raised approximately $4.7 billion with backing from Nvidia and Sequoia Capital. almakos.com · www.sbctv.gr · www.generation-nt.com
- [● 5 SOURCES] Beam uses three to four times less inference compute than competing models. cybernoz.com · www.sbctv.gr · futurumgroup.com · www.generation-nt.com · news.cnyes.com
- [● 6 SOURCES] The model was trained on 23.8 trillion tokens. almakos.com · www.prodiris.fr · www.sbctv.gr · futurumgroup.com · www.generation-nt.com · +1 more
- [○ 1 SOURCE] Reflection secured a compute deal with Nebius worth over $1 billion. cryptobriefing.com
- [● 9 SOURCES] The model activates 23 billion parameters per request to reduce inference costs. almakos.com · cybernoz.com · www.sbctv.gr · futurumgroup.com · www.watch.impress.co.jp · +4 more