YFarmX logoYFarmX

Reflection AI

Beam

A sparse model with an October open-weight release planned

Announced 5 October 20263 min readLarge Language Models

Editorial illustration of an expert-routing server blade, headed Beam, with Reflection branding and Nvidia green accents.

Key facts

501B
Total parameters
23B
Active per token
October 2026
Weights planned
Apache 2.0
Planned licence

Beam is an AI model for writing code, solving problems and using software tools. Reflection announced it on 5 October 2026, with downloadable weights planned later that month.

Listen to the model guide

Podcast: Beam's size and compute claims

Beam is an AI model for writing code, solving problems and taking actions through software tools. Nvidia-backed Reflection AI announced it on 5 October 2026, with an open-weight release planned for October.

An open-weight release provides the model’s learned numerical settings so developers can run and adapt it. Early access to Beam currently opens through Reflection’s waitlist.

How does a mixture of experts work?

A mixture-of-experts model routes each token, a piece of text, through selected parts of its network. Nvidia’s explanation describes a learned router that activates a subset of expert subnetworks for each token.

Beam has 501 billion total parameters and 23 billion active per token. That split lets a large model use a smaller share of its network at each step.

The full set of weights still needs storage and memory. Deployment also has to handle routing and communication between processors. For a team planning to host a model, active parameters help explain computation; total parameters help explain the size of the system it has to accommodate.

Schematic token routing through selected expert blocks, showing Beam's 501 billion total and 23 billion active parameters.
Selected experts process each token. The illustrated blocks are schematic. Sources: Reflection and Nvidia.

How does Beam compare with GLM-5.2?

Reflection’s launch table reports the following selected scores. These are the company’s published comparisons; each row uses a different evaluation.

Evaluation Beam GLM-5.2
DeepSWE v1.1 44.4 44.0
Terminal Bench v2.1 80.1 81.0
GPQA Diamond 90.5 91.2

Reflection claims comparable advanced-reasoning performance with three to four times less estimated inference compute. Its estimate multiplies active parameters by generated tokens and two operations per multiply-add. The calculation covers generation’s forward pass; prompt processing, context-dependent attention and serving overhead sit outside that estimate.

For developers comparing models, the useful next step is a repeatable test on their own codebase, tools and response-length settings. Record success rates, elapsed time, token use and the resources consumed by the complete task.

Reflection's approximate generation compute formula illustrated through active parameters, generated tokens and multiply-add operations.
Reflection estimates generation compute from active parameters and output tokens. The complete serving workload includes additional operations.

What will the Apache 2.0 release allow?

Reflection plans to publish Beam’s weights under Apache 2.0 later in October 2026. That is the announced release timetable.

The Apache licence permits use, modification and distribution, including commercial use, subject to its conditions. Redistribution includes providing the licence, retaining required notices and marking changed files. The licence also contains a patent grant and warranty terms.

Teams choosing an open-weight model can control their hosting environment and adapt the weights. They take responsibility for the deployment, its resources and how its tools access other systems.

Who is behind Reflection?

Reflection was founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou. Reuters reported in October 2025 that the Nvidia-backed company raised $2 billion at an $8 billion valuation. Those figures describe that funding round.

Beam’s announcement gives developers a model to assess for coding and tool-based work. The scheduled weight release will give hosting teams the files they need to test those claims on their own systems.