Model update · Official demonstration

Beam weights are open: from a game to tools people can revise

Continue from the astronaut game: Beam 1.0 weights and a 1M API are available. A dictation app shows how model-written code, later repairs and local speech recognition become source others can continue building.

The preview is dated October 5; its demonstration and API record retain the October 6 check. Weights were released October 16. The new access and dictation-tool sections were checked October 17, Beijing time.

Reflection's Beam game demo shows an astronaut, Earth and approaching asteroids.
An original frame at about three seconds in Reflection's demonstration. The announcement shows the animation. · Open full-size image
In this articleContents5

中文

Start with the game it produced

Reflection's October 5 Beam preview includes a p5.js astronaut game: dodge asteroids or fire beams. Its demonstration turns a coding request into a visible artifact.

Beam takes text. Generating graphics through code differs from reading a photograph. Kimi's SVG and image-description example and GLM's 3D penthouse offer useful connections: inspect the artifact, then follow the code, tools and human revisions.

Sources and further reading

中文

What do capability and compute each tell us?

The Beam preview is a sparse mixture-of-experts model with 501B total and about 23B active parameters. Size and active computation do not replace task outcomes. Preview launch scores are vendor-reported; our October 6 record had no independent Beam score. The new section covers the later 1.0 result.

The efficiency chart estimates generation compute, excluding prompt processing, attention and serving overhead. Keep that estimate distinct from task success, speed and actual charges; it is not an API bill.

Sources and further reading

中文

How far had preview access opened on October 6?

When checked October 6, documentation listed Beam-501B-A23B with 256K context counting input and output together, 128K maximum output and medium default effort. The beta API then opened gradually through a waitlist. The announcement's 1M effective training context was distinct from that early serving window.

The October 5 announcement planned Apache 2.0 weights and further artifacts later that month; GLM-5.3 and Kimi K3 already listed downloadable files. This table preserves that preview stage. The next section follows Beam's actual October 16 release.

Access stages in public materials checked October 6; use each provider's eligibility and license details.
ModelPublic materialWeight stage
Beam previewOfficial demos and beta API documentationPlanned for this month
GLM-5.3Model explanation, evaluations and deployment materialsFiles in the official repository
Kimi K3Model explanation, evaluations and deployment materialsFiles in the official repository
Sources and further reading

中文

Weights are open: what changes for access?

On October 16, Reflection released Beam 1.0 with BF16, FP8 and NVFP4 weights, a DFlash2 drafter, model cards and a safety technical report. Apache 2.0 main weights are now real files for deployment, modification or further training. The earlier preview's planned release has happened. Public sources do not establish that 1.0 and the October 5 preview are the same checkpoint, so their identities remain separate.

It remains a text-input, text-output model. The current Beam-1.0-501B-A23B API allows 1M tokens for input and output combined, up to 128K output and five effort settings; query the actual default for your organization. That does not rewrite the October 6 record of a 256K service. For self-hosting, 23B active parameters do not mean storing only 23B weights. Hugging Face estimates about 1 TB for BF16 or 280 GB for NVFP4 weights, before context-cache and runtime overhead.

There is also a new independent result: Artificial Analysis lists Beam 1.0 at 30 on October 17, without naming an effort setting. This is its text intelligence index, separate from the vendor's coding, tool and safety evaluations; it is not assigned retroactively to the preview. Open weights make inspection, adaptation and self-hosting possible. Success in a particular job still needs evidence from that job.

Access routes checked October 17; hardware recipes are examples, not minimum requirements or our device tests.
RouteWhat it enablesConditions
Hosted APIConnect coding tools or your serviceAccount and plan access; 1M capacity need not be filled on every request
Self-host weightsChoose precision, adapt or train furtherNVFP4 needs compatible hardware; the SGLang recipe uses four B300s, with FP8 for older GPUs
Sources and further reading

中文

From a game to a dictation tool people can revise

The accompanying Hugging Face post offers another concrete result: BeamSpeak, a menu-bar dictation tool for Apple Silicon Macs. Command+Shift+Space starts and stops recording, and the transcript can reach the clipboard. Local recognition uses Parakeet v3 through MLX. Beam's role is writing the app's code; this is not evidence of the 501B model running on that Mac.

The build notes preserve the handoff: Beam generated the first implementation through Pi; after a provider rate limit, Codex reviewed and repaired it and renamed it BeamSpeak. The author separately checked local recognition, microphone capture, the physical shortcut and clipboard delivery. Automatic paste remains unverified. The video shows the recording pill and text; the notes distinguish what that view establishes from what still needs checking.

What is public is source code with development instructions. It needs Xcode command-line tools, Python 3.13, uv, ffmpeg and the speech model. The development bundle is not notarized and cannot be moved as a standalone .app. It is a readable, repairable starting point for a builder; someone seeking an install-and-go dictation app should first check these conditions. We read the notes and inspected video frames, without installing the app, calling Beam or downloading model weights.

An original BeamSpeak video frame: the floating dictation pill reads Ready, and English text appears in a browser search field.
A frame at about 13 seconds in the author's video. The dictation pill is below an English query; the running notes still mark automatic paste unverified, so this frame alone does not establish that step. · Open full-size image
Sources and further reading