Reflection Beam: Open 501B Model and Choice

Reflection released Beam, an open model with 501 billion parameters, 23 billion of which are active for each token (501B-A23B architecture). It was trained from scratch in the United States.

Beam does not rank first in benchmarks.

For a company choosing a model, its ranking matters less than replacement time: can the team switch models within a week when the leader changes? Reflection launched more than a year ago, positioning itself around strong coding and its own approach to reinforcement learning (RL).

Throughout this time, the company operated in stealth mode, longer than Thinking Machines.

It was unclear whether the model would ever be released, while open models had been launching continuously throughout the year. Beam is built on a Mixture of Experts (MoE) architecture.

The model stores 501 billion parameters but activates only 23 billion “experts” suited to the task for each token.

Memory is needed for all 501 billion parameters, while the computation per token corresponds to a model several times smaller.

That is why MoE models are cheaper to infer than dense models of the same size, provided the hardware can hold the full set of weights.

A resilient solution looks simple but is built on engineering discipline.

A layer between applications and models handles routing, policies, and logging. In KT.Team solutions, this role is performed by the LLM & Security Gateway: the application connects to one interface, behind which you can run Anthropic Claude, Qwen, Llama, GigaChat, or open weights deployed in your environment.

The gateway enforces policies: which data can leave the system, who made how many requests, and what appeared in the response. The second layer is data access.

If corporate knowledge is connected through RAG and MCP servers rather than embedded in prompts for a specific model, switching models comes down to configuration and regression testing.

We build such environments as part of AI-native integration: we select 20–50 real customer requests, run candidate models on them, and base the choice on the results for those requests.

A hypothesis to test on your own tasks: in common scenarios such as request classification, document-field extraction, or product-card generation, the performance gap between the first- and fifth-ranked models may be smaller than the difference in deployment cost.

Benchmarks do not prove this, so without testing on your own data, the choice remains a guess.

Beam currently lags behind the best open models.

However, it expands the list of viable options for those who consider the origin of the weights critical.

Those who do not choose it benefit too: competition among open models reduces costs and timelines.

It is wiser to invest in a layer that lets you change models without interrupting the product than to bet on a specific model.

The ranking leader will change, while the gateway, policies, and test query set will continue to work with any model.