Reflection launches Beam, a 501B-parameter open-weight model it says beats Chinese rivals on cost

Reflection AI unveiled Beam on Monday, a 501-billion-parameter open-weight model that it claims matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. The launch makes the two-year-old startup the most serious Western challenger yet to the open-model dominance of DeepSeek, Qwen, and Z.ai.

Beam is a text-only mixture-of-experts model with 23 billion active parameters. It was pre-trained on 23.8 trillion tokens and supports a 1 million token context window. For comparison, Z.ai's GLM 5.2 has roughly 744 billion total parameters with 40 billion active — meaning Beam activates a smaller slice of its network per query, which cuts the cost of running it.

The company laid out the details in a lengthy blog post Monday, describing Beam as a model trained on high-compute reinforcement learning to be effective at reasoning, coding, and agentic tasks at "a fraction of the token cost and inference time compute" of rivals. Reflection calls it a "workhorse model" for enterprises, the public sector, and developers.

Why Beam matters for the open-model race

The stakes here are straightforward. Chinese labs currently set the pace for open-weight frontier models, and Western developers who want to fine-tune, self-host, or audit a top-tier model have had few options outside DeepSeek, Qwen, and Z.ai. Reflection is explicitly positioning Beam against all three — and against closed labs like Anthropic and OpenAI, plus Western open-model players Mistral, Meta, and Cohere.

Reflection says Beam scores on par with Z.ai's GLM-5.2 on advanced reasoning benchmarks and outperforms today's leading Western open models while using "3-4x less inference compute." Those claims have not been independently verified, and the company did not respond in time to TechCrunch's requests for more information.

Beam's most direct U.S. rival may be Inkling, the open model from Mira Murati's Thinking Machines Lab released in July. Reflection's own benchmarks show Beam outscores Inkling on four coding tests where both report results. The comparison comes with a caveat: Inkling is multimodal, while Beam handles text only.

What are the model's specifications?

The headline numbers from Reflection's announcement:

• 501 billion total parameters, 23 billion active per query (mixture-of-experts architecture)
• 23.8 trillion tokens of pre-training data
• 1 million token context window
• Text-only, tuned via high-compute reinforcement learning for reasoning, coding, and agentic tasks
• Weights and full technical details to be released this month

For scale, Z.ai's GLM 5.2 runs roughly 744 billion total parameters with 40 billion active. Reflection's bet is that a leaner activation footprint translates directly into lower serving costs — the argument that made DeepSeek's models disruptive when they arrived.

The launch confirms reporting from Axios over the weekend that the startup was close to shipping. Reflection's Monday blog post marked the official debut.

Who is behind Reflection — and how much compute do they have?

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, per PitchBook. Its last round valued the company at a $25 billion pre-money valuation — an aggressive number for a startup that, until this week, had no public model.

Compute, not capital, is typically the binding constraint for frontier training. Reflection has moved to lock it down. This summer, the startup signed deals collectively worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia's GB300 chips through 2029. That supply agreement gives Reflection a multi-year runway to train successors to Beam and compete for customers weighing a move away from Anthropic's and OpenAI's closed models or the cheaper open-weight alternatives from Chinese labs.

What is an "AI factory," and who is buying?

Beam is not the end product. Reflection is aiming the model, and future ones, at enterprises and sovereign nations with a pitch to build "AI factories" — systems that let institutions train Reflection's models on their own proprietary data and run customized AI locally, on infrastructure they control.

Nvidia CEO Jensen Huang, whose company backs Reflection, has long championed the "AI factory" idea and pushed to strengthen the open AI ecosystem. The vision aligns neatly with Nvidia's business: those local systems would run on Nvidia GPUs, including the GB300 chips Reflection has already reserved through 2029.

There are early signs of demand. Axios reported that hedge funds and trading firms are among those eager to build such systems — institutions with proprietary data and strict requirements about where that data can travel. Reflection has already begun testing the sovereign AI factory concept with Shinsegae Group in South Korea.

The sovereign angle is where policy stakes come in. Governments increasingly want frontier AI capability without depending on U.S. closed labs or Chinese open models. An American open-weight model that enterprises and states can run on their own hardware serves both procurement requirements and data-sovereignty rules — a market that DeepSeek's and Qwen's Chinese origin has complicated for Western buyers.

When can developers get the weights?

Reflection says it will release Beam's weights and full technical details this month. Distribution will run through hyperscalers and neoclouds, with integrations across open source libraries at launch.

That timeline matters because Reflection's claims — the GLM-5.2 parity, the 3-4x inference-compute advantage, the coding wins over Inkling — remain self-reported until outside evaluators can run the weights. The open-source community's benchmarking infrastructure typically delivers independent verdicts within days of a major open-weight release.

If the numbers hold, Beam gives Western enterprises a cost-competitive, self-hostable alternative to Chinese open models for the first time, and puts pricing pressure on the closed labs at the top of the market. If they don't, Reflection's $25 billion valuation rests on a model that will face immediate, public scrutiny — the standard bargain of open-weight releases, and one reason the next few weeks of independent benchmarking will do more to establish Beam's position than Monday's blog post.