Beam’s weights are now available. The practical question has moved from a promised release to the checkpoint, infrastructure, and access path you can verify.
October 19 update: replace the preview assumption
Reflection’s October 16 release makes Beam weights available. The official model card lists Apache 2.0 licensing, 501 billion total parameters, and 23 billion active parameters.
The text-input, text-output model supports a context window of up to one million tokens. Published artifacts include BF16, FP8, and NVFP4 versions, with a separate drafter artifact for supported serving configurations.
Downloadable weights still need a serving plan
The active parameter count describes computation, not the memory footprint of the entire model. Choose a checkpoint format and serving recipe before estimating hardware, concurrency, or long-context capacity.
Keep the license, model revision, inference engine, and relevant configuration together. A different numerical format or chat template can change behavior enough to matter in a comparison.
Ollama’s offer runs in the cloud
Ollama’s Beam announcement describes a cloud offering that provides temporary free access and zero data retention. Those are provider terms, not a permanent price commitment or proof that the model runs locally on an ordinary laptop.
No fixed end date for the introductory offer was verified in this review. Check the current service terms before budgeting around it, particularly if the workflow depends on continuing no-cost access.
Separate safety documentation from the full report
The model card links a safety report while still marking the full technical report as forthcoming. Those are different documents. Neither a report’s existence nor a vendor benchmark result establishes how the model handles your own permissions and data.
I would begin with a bounded coding task in a disposable repository. Give the reviewer the original issue, the allowed files, and the expected checks.
- Record whether the proposed change solves the issue.
- Inspect the patch for unrelated edits.
- Run the relevant checks and review their actual output.
- Track failed attempts and human repair time.
This update corrects the earlier preview framing. The proposed evaluation above has not been run as a hands-on Beam test for this article.
Compare the Step 5 artifact correction and Muse Spark Open license review before choosing another large-weight deployment.
Read the sources and plan your next step
- Reflection release announcement
- Model card, license, and available formats
- Ollama cloud offer and data-retention statement
Will your first evaluation use hosted access or a specified self-hosted configuration?
Checked October 19, 2026. This is source-based reporting and analysis; proposed checks are not hands-on test results.
