Reflection AI has introduced Beam, its first frontier open-weight model, with an argument aimed squarely at the economics of deploying advanced artificial intelligence. TechCrunch reported on October 5 that the Brooklyn startup is presenting Beam as a Western competitor to leading Chinese open models, including systems from DeepSeek, Qwen and Z.ai. The company says the model can deliver comparable reasoning performance while demanding substantially less compute at inference time. That performance claim has not been independently verified.

According to TechCrunch, Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. Reflection says it pretrained the system on 23.8 trillion tokens and equipped it with a one-million-token context window. The company describes the model as suited to reasoning, coding and agentic work, and calls it a general-purpose workhorse for enterprises, public-sector organizations and developers.

A large translucent AI model contains a smaller active engine amid token streams.
Reflection says Beam has 501 billion total parameters but activates 23 billion at a time.

The central sales pitch is efficiency rather than size alone. Reflection says Beam performs on par with Z.ai’s GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. TechCrunch noted that GLM-5.2 has roughly 744 billion total parameters and 40 billion active parameters. Reflection also says Beam beats leading Western open models on its reported tests, though those results still depend on the company’s own benchmark disclosures.

That caveat matters because the release is entering an increasingly crowded contest in which labs choose different model formats, evaluation sets and reporting methods. TechCrunch reported that Reflection’s comparisons show Beam ahead of Thinking Machines Lab’s Inkling on four coding tests for which both organizations publish results. The comparison is not complete, however: Inkling is multimodal, while Beam handles text only, making a simple overall ranking difficult.

Reflection is placing the model between two competitive camps. On one side are closed providers such as OpenAI and Anthropic, whose flagship systems are accessed as services. On the other are open-weight developers in China and Western companies including Mistral, Meta and Cohere. By offering model weights, Reflection is trying to make customization and local deployment part of the product rather than an afterthought.

A local AI facility connects securely to private data vaults and cloud infrastructure.
Reflection’s broader plan is to help institutions build customized local AI systems.

The company’s larger plan extends beyond distributing a downloadable model. TechCrunch reported that Reflection wants to sell what it calls AI factories to enterprises and sovereign nations. Under that concept, customers would use Reflection models and their own proprietary data to build customized systems that can operate locally. The company has already begun testing the sovereign version of that approach through a partnership with South Korea’s Shinsegae Group, according to the report.

Reflection has raised roughly $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners, TechCrunch reported, citing PitchBook. Its most recent financing valued the company at $25 billion before the investment. The startup has also committed heavily to infrastructure: deals with SpaceX and Nebius worth more than $7 billion in total are intended to provide access to Nvidia GB300 chips through 2029. Those commitments show that an efficiency-focused model strategy still relies on large amounts of training capacity.

The immediate test will come when outsiders can inspect and run the system. Reflection says it plans to release Beam’s weights and full technical details during October, with distribution through hyperscale cloud providers and specialized AI clouds, plus integrations with open-source libraries. Until those materials arrive and independent evaluations are available, Beam’s reported cost and performance advantages should be treated as company claims rather than established results.