Beam has 501 billion total parameters, but Reflection AI says it activates 23 billion per token. That distinction helps explain the compute needed to generate each token: the 23B figure is more relevant to an estimate of token-generation compute, while 501B describes the model’s full parameter scale. Neither number alone tells you how fast Beam will respond or what it will cost to run.
What “23B active parameters” means
Beam is a sparse Mixture-of-Experts (MoE) model, according to Reflection AI’s October 5, 2026 announcement. In an MoE model, the full set of parameters is not necessarily used for every token. Reflection reports 501 billion parameters in total and 23 billion active per token.
As an Amazon Associate I earn from qualifying purchases.
Think of 501B as the model’s overall parameter scale and 23B as the portion involved in processing a given token. The active count helps explain the compute involved in token generation, but it does not make Beam a 23B-parameter model. Nor does it alone establish how much memory a deployment needs: the announcement does not give minimum local hardware specifications.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsWhy the active count matters—and what it cannot tell you
A dense model of comparable total size generally uses its parameters differently from a sparse MoE model. By activating a subset of parameters for each token, an MoE model can require less generation compute per token than a dense model with a similar total parameter count. That is why the active count is useful when discussing compute efficiency.
Reflection says its comparison estimates generation forward-pass compute at roughly two times the number of active parameters multiplied by the mean generated tokens per attempt. It counts both reasoning and final-answer tokens. For MoE models, Reflection says it used parameters activated per token rather than total model size.
Using that method, Reflection estimates Beam uses 3–4× less inference compute than GLM-5.2 on comparable advanced-reasoning tasks. This is a company-published estimate, not a measured reduction in end-to-end serving cost, response time, or power use. Reflection’s calculation excludes prompt prefill, context-dependent attention operations, and serving overhead. Those factors can affect actual deployment performance and cost.
What Reflection says Beam is built to do
Reflection calls Beam its first open-weight model and describes it as designed for coding, reasoning, and agentic workloads. Its announcement reports 23.8 trillion pretraining tokens, more than 100 million reinforcement-learning rollouts over four weeks on 10.5K NVIDIA GB300 GPUs, and approximately 1.3 billion sandboxes used for training and grading. It says the high-compute reinforcement-learning run used a maximum context length of 256K tokens, while midtraining extended Beam’s effective context length to 1M tokens.
Free tools Windows power users keep installed
One-click scans. No signup required.
These are figures reported by Reflection, not independently audited measurements established by the announcement. The model is text-only, according to the company; it can work with other modalities when their information is represented as text.
How Beam’s published benchmark results compare
Reflection’s company-published table shows Beam performing strongly on some tasks and trailing listed competitors on others. The scores below are the values included in the announcement; an unlisted result is omitted rather than treated as a zero.
| Benchmark | Beam | Other listed scores |
|---|---|---|
| SWE-bench Verified | 80.9 | Inkling 77.6; Nemotron 3 Ultra 70.7 |
| Terminal Bench v2.1 | 80.1 | GLM 5.2 81.0; GLM 5.3 88.2; Kimi K3 88.3; Qwen 3.8 Max 86.6; DeepSeek V4.1 Flash 90.6 |
| AIME 2026 | 97.8 | Inkling 97.1; GLM 5.2 99.2 |
| GPQA Diamond | 90.5 | GLM 5.2 91.2; GLM 5.3 91.7; Kimi K3 93.5; Qwen 3.8 Max 92.6; DeepSeek V4.1 Flash 90.9 |
Reflection characterizes Beam as competitive with GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks, while acknowledging Kimi K3 remains ahead on raw capability. Those are the company’s assessments of its published results, not an independent ranking. A benchmark score and an estimated compute figure answer different questions: a model can score lower on a task while using less estimated generation compute.
Rank #4
For a useful comparison, check the task and benchmark protocol, token-generation settings, context and prefill treatment, and whether results were independently reproduced. Do not treat a benchmark table and a compute estimate as a single overall ranking.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Availability and evidence limits as of the announcement
On October 5, 2026, Reflection said Beam was undergoing final red-teaming and evaluations and was available to select users through an early-access waitlist. It planned to release weights under an Apache 2.0 license, alongside a technical report, model card, documentation, and developer materials later in October. Those were plans announced on that date, not confirmation that public weights were already available.
Reflection also said safety evaluation results would be published with the future technical report. The announcement itself does not provide those results or establish independent safety validation. Its benchmark scores, training figures, compute estimate, and release plans should therefore be read as company-reported claims.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
