Reflection has published a technical report on Beam, a sparse mixture-of-experts model with 501.3 billion total parameters and 23.4 billion active parameters. The report describes training on 23.8 trillion tokens and a one-million-token context window, giving developers a substantial new set of model specifications to examine.
Reflection AI Watch analysis
What happened
The report says Beam is designed for coding, reasoning and agentic workflows. Reflection describes a two-stage midtraining phase and a large-scale asynchronous reinforcement-learning campaign as part of its training process. The source is the company’s own technical report, published on 13 November.
Those details are useful, but they are not a performance leaderboard: the supplied report excerpt gives no benchmark scores or independent comparisons. Nor does it establish that model weights are available, so this is a technical disclosure, not evidence of a model developers can download and run.
Why it matters
A one-million-token context window and a model this large are consequential specifications for teams assessing long-context and agentic systems. The active parameter count also gives a more informative picture of the sparse architecture than the headline total alone. Whether those design choices translate into useful results is the next question, not something the parameter count settles.
Our read
This deserves more than a passing mention: Reflection has put concrete architecture and training details into the public record. The report gives readers something specific to scrutinise, though the excerpt leaves key practical questions, including benchmark performance and access to the model, unanswered. Big numbers are a start; they are not a substitute for trying the thing.
What to watch
- Whether Reflection publishes benchmark results for coding, reasoning and agentic tasks.
- Whether model weights or an inference service become available, and on what terms.
- How the one-million-token context window performs on real workloads.
Discussion spark: When judging a model like Beam, should architecture and training details count as meaningful progress before independent benchmarks and practical access are available?
Sources and evidence
- Beam Model Technical Report: Architecture and Training of Reflection's Open MoE Model (13 November 2026, 06:19 UTC)
Independent WittyWires tracker for public updates about Reflection AI. Not affiliated with or endorsed by Reflection AI; this is not an official account.