Reflection Beam: Open-Weight Model Takes On Chinese Rivals
Reflection AI has released details of Beam, the first frontier open-weight model from the Brooklyn-based startup. The company says Beam matches the best Chinese open models on advanced reasoning benchmarks while using far less compute. If that holds up, it would add pressure to the race for a Western alternative to DeepSeek, Qwen and Z.ai.
The launch follows Axios reporting over the weekend that Reflection was close to shipping. On Monday the company followed up with a long blog post on how the model was built and what it is meant to do.
What Beam is
Beam is a text-only mixture-of-experts model. In this design, only part of the network is active for any given token, which keeps running costs down. Reflection trained it with high-compute reinforcement learning, aiming at three areas: reasoning, coding and agentic tasks. The company says Beam handles them at "a fraction of the token cost and inference time compute" of competing models.
The key specifications:
- Total parameters: 501 billion
- Active parameters: 23 billion
- Pretraining data: 23.8 trillion tokens
- Context window: 1 million tokens
For comparison, Z.ai's GLM-5.2 has about 744 billion total parameters, with 40 billion active.
The performance claims
Reflection says Beam performs on par with GLM-5.2 on advanced reasoning benchmarks and beats today's leading Western open models, while needing "3-4x less inference compute." None of these results have been independently verified yet.
The company describes Beam as a "workhorse model" for enterprises, the public sector and developers. That is a practical pitch rather than a claim to the top of every leaderboard.
Reflection is competing on several fronts at once. On one side are the closed labs, Anthropic and OpenAI. On another are the popular open models from Chinese developers, including DeepSeek. And there are Western open-model makers such as Mistral, Meta and Cohere.
Its closest U.S. competitor may be Inkling, the open model that Mira Murati's Thinking Machines Lab released in July. According to Reflection's own numbers, Beam scores higher than Inkling on four coding tests where both have published results. The comparison is not like for like, though. Inkling is multimodal, and Beam handles text only.
Money and compute
Reflection was founded in 2024 by two former Google DeepMind researchers. According to PitchBook, it has raised about $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners. Its most recent round set a pre-money valuation of $25 billion.
The startup has also been securing compute, which it needs to train frontier models that can pull customers away from closed systems and from cheaper Chinese open-weight options. This summer, Reflection signed deals with SpaceX and Nebius worth more than $7 billion combined. The agreements give it access to Nvidia's GB300 chips through 2029.
The "AI factory" pitch
Reflection's main targets for Beam and future models are enterprises and sovereign nations. The product idea is what it calls "AI factories": institutions train Reflection's models on their own proprietary data to build customized AI systems that run locally.
The concept is a familiar one from Nvidia CEO Jensen Huang, whose company backs Reflection. Huang has long promoted AI factories and a stronger open AI ecosystem. That vision also suits Nvidia, since its GPUs would power those systems.
According to Axios, hedge funds and trading firms are among the groups interested in building such systems. Reflection is already testing a sovereign AI factory partnership with Shinsegae Group, a South Korean retail and business conglomerate.
The company says it will publish Beam's weights and full technical details this month. At launch, the model will be distributed through hyperscalers and neoclouds (newer, GPU-focused cloud providers) and integrated into open source libraries. Reflection did not respond in time to TechCrunch's requests for further comment.
Our Take
Beam's significance depends on whether outside testing confirms Reflection's numbers. A model that matches GLM-5.2 while using three to four times less inference compute would be a real efficiency gain. For now, though, every figure comes from the company. Readers should treat the benchmarks as a starting point and wait for independent evaluations once the weights are public.
The timing fits a wider pattern. Chinese labs have set the pace in open-weight AI for some time, and Z.ai's GLM line has kept moving forward since GLM-5.2. Western companies are answering with open models of their own, from Thinking Machines' Inkling to regional efforts such as Aleph Alpha's Kolibri. Reflection is betting that efficiency, not just raw scores, will win business customers who care about running costs.
The AI factory strategy may matter more than the benchmarks. Selling to governments and regulated firms that want models on their own data and infrastructure is a different business from selling API access. With Nvidia as an investor and chip access secured through 2029, Reflection appears to have the resources to pursue it.
Three things are worth watching. First, whether independent tests support the compute-efficiency claims. Second, whether the Shinsegae pilot becomes a full deployment. Third, whether the lack of multimodal support limits Beam against rivals such as Inkling.