Reflection Beam: 501B-Parameter Open-Weight AI Model
Reflection's Beam packs 501 billion parameters with 23 billion active, and the startup claims it matches GLM-5.2 using three to four times less compute.
In short
Reflection has unveiled Beam, an open-weight language model with 501 billion parameters that activates just 23 billion per token and, by the company's own account, matches China's GLM-5.2 on reasoning benchmarks.
At a glance
- 501 billion total parameters, 23 billion active per token - a text-only mixture-of-experts design.
- Context window of 1 million tokens; pretrained on 23.8 trillion tokens from the public web and commercial sources.
- Beam Base trained on 6,144 GPUs in under four weeks; the RL stage used 10,000 GB300s and 1.3 billion sandboxes.
- Reference point GLM-5.2: roughly 744 billion parameters, 40 billion active.
- Weights, docs and fine-tuning tools are due in October 2026; access is early-access only for now.
Reflection has unveiled Beam, an open-weight language model with 501 billion parameters that activates just 23 billion per token and, by the company's own account, matches China's GLM-5.2 on reasoning benchmarks.
The pitch is efficiency, not the leaderboard
Reflection is not claiming the best model in the world. It is claiming parity with a leading Chinese open model at three to four times less inference compute. On four coding tests, the company says Beam comes out ahead of Inkling, the open model that Mira Murati's Thinking Machines Lab released in July.
TechCrunch puts GLM-5.2 at roughly 744 billion parameters with 40 billion active. SiliconANGLE frames the same gap differently: GLM-5.2 carries about 250 billion more parameters, and Beam handles some jobs on a third to a quarter of the hardware.
Inside the model
Beam is a text-only mixture-of-experts system with a context window of 1 million tokens. Pretraining ran over 23.8 trillion tokens drawn from the public web and commercial sources. Neither outlet published raw benchmark scores, so the efficiency claim currently rests on the vendor's own framing.
By SiliconANGLE's account, Beam Base came together on a 6,144-GPU cluster in under four weeks. A reinforcement-learning stage then spun up 10,000 GB300 cards running 1.3 billion sandboxes, and took another four weeks. The result approaches Qwen 3.8-Max, a model with more than 2 trillion parameters, but trails frontier systems such as Anthropic's Claude Fable 5.1.
Who Reflection wants to sell to
The target buyers are enterprises and sovereign states rather than consumers. Reflection wants to sell „AI factories“ - packages that let an institution train the company's models on proprietary data and run the result locally. A first test of that idea is under way with Shinsegae Group in South Korea.
The balance sheet behind it is unusually large for a two-year-old lab. Founded in 2024 by two former Google DeepMind researchers, Reflection has raised about $4.7 billion and was last valued at $25 billion pre-money, with Nvidia, Sequoia Capital and Lightspeed Venture Partners among its backers. Over the summer it signed deals worth more than $7 billion combined with SpaceX and Nebius for access to Nvidia GB300 chips through 2029, and SiliconANGLE reports the SpaceX portion alone at $6.3 billion for rented GB300 NVL72 appliances.
What is not confirmed yet
Beam is currently limited to an early-access program. Weights, documentation and fine-tuning tools are due later in October 2026, distributed through hyperscalers and neoclouds and integrated into open source libraries at launch. Neither source names a license, so „open weight“ remains the company's own label rather than a published set of terms.
Every performance figure above comes from Reflection, and no independent evaluation exists so far. TechCrunch notes that the company did not answer its questions in time. A third trade publication we tried to consult was unreachable during reporting, so details such as the model's expert count could not be cross-checked.
FAQ
What is Reflection Beam?
Beam is an open-weight language model from the U.S. startup Reflection, built as a mixture of experts with 501 billion total parameters and 23 billion active per token. It is text-only and handles a context window of 1 million tokens.
When will the Beam weights be released?
Reflection says weights, documentation and fine-tuning tools arrive in October 2026, shipped through hyperscalers and neoclouds. Until then the model is limited to an early-access program, and the company has not named the license that will apply.
Is Beam better than GLM-5.2?
Reflection claims comparable reasoning quality at three to four times less inference compute, plus a lead over Inkling on four coding tests. Those are vendor figures: no independent benchmark results have been published.
