Reflection Beam
Beam is Reflection’s open-weight text model for coding, reasoning and long tool workflows. Its official card describes 501.3B total parameters, 23.4B active per token and a 1M-token context window. Downloadable checkpoints and hosted access have different operational requirements.
Inspect the Beam model contract ↗Why it matters
The weight release gives teams a concrete checkpoint to evaluate. Sparse activation reduces computation per token; it does not reduce the resident weights to 23.4B parameters.
Choose a checkpoint before choosing a server
The official repository publishes BF16, FP8 and NVFP4 variants under Apache 2.0. Match the numerical format to an implementation that supports its architecture and tool/reasoning parsers. Preserve the model and tokenizer revision when comparing outputs.
- Official checkpoint and license
Inspect the selected numerical format and model card.
Hosted access and self-hosting solve different problems
Hosted access avoids provisioning a large model, while self-hosting gives you operational control and the responsibility for memory, runtime, concurrency and logging. Reflection’s developer guide is the place to verify the current API contract; no universal token price or successful request is asserted here.
- Developer documentation
Check current authentication, model naming and access.
Evaluate an agent workflow with a reference outcome
Start with a small repository issue, a known test suite and a fixed tool budget. Record the checkpoint, harness, reasoning settings, failed actions and final tests. A text model can request tools, but execution and permissions belong to the agent host. This is a proposed evaluation method, not a reported Beam result.
Read efficiency claims with their conditions
Reflection and Google Cloud publish efficiency and serving results for selected configurations. Treat these as attributed measurements with hardware, parallelism and workload conditions. They do not establish a universal speedup, a laptop deployment or better quality on your repository.
- Google Cloud deployment discussion
Read the infrastructure and measurement conditions.
Useful first jobs
- Choose a checkpoint before choosing a server
- Hosted access and self-hosting solve different problems
- Evaluate an agent workflow with a reference outcome
Frequently asked questions
Is Beam an image model?
No. The released Beam card specifies text input and text output.
Can I run Beam as a 23B model?
No. The active parameter count describes computation per token; the checkpoint still contains about 501B total parameters.
Sources and evidence
Checked October 21, 2026 against the linked primary sources. Specifications and examples are publisher documentation; no independent model or hardware run is claimed.
- October 16 weight release ↗
Reflection · Primary source
- Checkpoint, numerics and model specifications ↗
Reflection · Primary source
- Hosted access documentation ↗
Reflection · Primary source