CoAI
News

Reflection AI Releases Weights of Beam, a 501-Billion-Parameter Model for Programming and Agents, Under Apache 2.0

On October 16, 2026, the U.S. AI startup Reflection AI released on Hugging Face the weights of Beam, the AI model it announced on October 5. Of its 501 billion total parameters, it uses 23 billion in each pass, and it is licensed under the permissive Apache 2.0. The API is available as a beta with a free trial allowance.

October 16, 20264 min read

On October 16, 2026, the U.S. AI startup Reflection AI released the weights (the data that makes up the trained model) of its AI model Beam. Beam, announced on October 5, is built for programming, reasoning and "agent" use, in which AI works through tasks while using tools. The weights can be downloaded from Hugging Face under Apache 2.0, a license that allows free use, including commercial use.

Reflection AI and Beam

Reflection AI is a New York-based startup founded in March 2024 by Misha Laskin, who worked on Gemini development at Google DeepMind, and Ioannis Antonoglou, a co-creator of the Go-playing AI AlphaGo. It has made developing U.S.-made openly released models (open-weight models) its mission, and in October 2025 it raised US$2 billion at a valuation of US$8 billion. Beam is the company's first open-weight model and is positioned as the first in a series.

According to the official introduction page, Beam's main specifications are as follows.

  • Design: MoE (Mixture of Experts). It has 501 billion (501B) parameters in total, of which 23 billion (23B) are used in each pass
  • Amount of training data: 23.8 trillion tokens
  • Usable input length (context): 1 million tokens

MoE is a design in which, even though the model as a whole is large, only some "expert" parts run for each pass, so it can run fast and cheaply for its size.

Released weights and license

The weights have been released on Hugging Face in the following formats.

  • BF16: the original-precision format
  • FP8, NVFP4: formats that shrink the data to reduce the memory required
  • DFlash2: a helper model for faster text generation (used with a method called speculative decoding)

Apache 2.0 is a widely used open-source license that permits commercial use, modification and redistribution without restrictions based on revenue or number of users. This stands in contrast to Muse Spark Open, released by Meta the day before on October 15, which came with its own license that places conditions on offerings by large companies.

How to use it

Beam can be used in the following ways.

  • API: offered as a beta with a free trial allowance on Reflection AI's developer site (platform.reflection.ai)
  • Cloud and other platforms: available on AWS, Google Cloud, Microsoft Foundry, Together AI, Ollama and others (more than 40 partners and integrations)
  • On-premises servers: can run on Dell and HPE servers

API pricing has not been announced. There is no specific information about use from Japan or support for Japanese.

Evaluation results

The main figures on the official introduction page are as follows. Beam's figures were published by Reflection AI, and the company says it used data from the evaluation sites Artificial Analysis and DataCurve for the figures of the other companies' models it compared against. There has been no third-party verification yet.

  • SWE Bench Pro v2-Hard (difficult software development tasks): 77.2
  • Terminal Bench v2.1 (command-line operations): 80.1
  • AIME 2026 (math competition problems): 97.8
  • GPQA Diamond (graduate-level science questions): 90.5

For Meta's Muse Spark Open, released the day before, Meta reports 76.4 on Terminal-Bench 2.1 (measured with an agent called Terminus 2) and 89.0 on GPQA Diamond. However, the measurement conditions are not necessarily the same, so the figures cannot be compared directly.

CoAI's take

Among large open models released by U.S. companies, Beam has especially few license restrictions. Under Apache 2.0, even companies with large revenues can build it into their own services without a separate agreement. In the past week, Meta, China's StepFun and Reflection AI have released large open models one after another, and the options for AI you can run in your own environment have suddenly expanded.

On the other hand, the published performance figures come from Reflection AI itself, and third-party evaluations are still to come. With 501 billion parameters in total, the model is large, and running it in-house requires servers with big GPUs. A good first step is to use the free API trial allowance to see whether it suits your own work.

Sources

Back to news