A Brooklyn startup just put its name in the race for the best open AI model. TechCrunch reports that Reflection AI has unveiled Beam, its first frontier open-weight model. The two-year-old company claims Beam matches the leading Chinese open models on advanced reasoning tests while costing far less to run.
First, what does open-weight mean? The weights are the numbers a model learns during training. When a company releases them, anyone can download the model and run it on their own hardware instead of renting it through a chatbot website. That is why open models from Chinese labs like DeepSeek, Qwen, and Z.ai have become so popular with developers and businesses.
Reflection shared the details in a long blog post Monday, confirming earlier reporting from Axios that a launch was close. Beam is a text-only mixture-of-experts model. That means only part of the model wakes up for each request, which saves compute. It has 501 billion parameters in total, with 23 billion active at a time. It was pretrained on 23.8 trillion tokens and has a 1 million token context window. For comparison, TechCrunch notes that Z.ai's GLM-5.2 has roughly 744 billion total parameters with 40 billion active.
The company says Beam was trained with high-compute reinforcement learning to be good at reasoning, coding, and agentic tasks, at "a fraction of the token cost and inference time compute" of rivals. On advanced reasoning benchmarks, Reflection says Beam scores on par with GLM-5.2 and beats today's leading Western open models while using "3-4x less inference compute." These claims have not been independently verified yet, and I would wait for outside testing before taking them as settled.
Reflection is picking fights on several fronts. It is going after closed labs like Anthropic and OpenAI, the Chinese open models, and Western players like Mistral, Meta, and Cohere. TechCrunch says its closest U.S. rival may be Inkling, the open model from Mira Murati's Thinking Machines Lab that came out in July. Reflection's own numbers show Beam ahead of Inkling on four coding tests where both report results. One caveat is that Inkling handles images and other media, and Beam only does text.
There is real money behind this. Reflection was founded in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, according to PitchBook. Its last round valued it at $25 billion before the new money. This summer it signed deals worth more than $7 billion with SpaceX and Nebius to get access to Nvidia's GB300 chips through 2029.
The business plan is what Reflection calls "AI factories." Institutions would train Reflection's models on their own private data to build a custom AI system they run locally. Axios reported that hedge funds and trading firms are interested, and Reflection is already testing a sovereign AI factory partnership with Shinsegae Group in South Korea. The company says it will release Beam's weights and full technical details this month.
Why should a regular person care? Because cheaper models change who gets to build with AI. If a model really needs three to four times less compute to answer a question, that cost savings can show up as cheaper apps, faster responses, and tools that run inside a company instead of sending your data to someone else's servers.
I like seeing more competition in open models. Every serious option that ships its weights gives developers like me more freedom to pick what fits, test it ourselves, and switch when something better shows up. The real test comes when the weights land this month and people outside Reflection start running it.
