AI Trend Notifier
EN한
← wiki

$ cat wiki/models/beam-1-0.md

Beam 1.0

modelupdated 2026-10-18created 2026-10-18

Compared with

Spec

AttributeValue
DeveloperReflection
Released2026-10-16
Announcedunknown
Context window1M tokens
Pricingunknown
LicenseApache 2.0
AvailabilityOpen weights; Reflection API public beta; partner endpoints
Release and license: (source) Context and modality: (source) The reviewed release does not state a token price; an earlier announcement date is not established by these two captures.

Release Date

Weights became available on 2026-10-16, in BF16, FP8 and NVFP4, with a speculative-decoding drafter. The model is text-only, with 501.3B total / 23.4B active parameters according to the model card. (source) (source)

Benchmarks

Reflection's final model card reports the following results. These are vendor evaluations, with temp=0.7 and top_p=0.99, not independent measurements. (source)

Row labelBeam (BF16)Beam (FP8)Beam (NVFP4)
DeepSWE v1.143.3%43.0%42.7%
Terminal Bench v2.180.1%78.9%78.4%
MCP Atlas80.8%81.5%80.5%
The columns identify numerical formats, not reasoning-effort settings. A smaller representation does not improve every benchmark. (source)

Use Cases

Coding, tool use and long-horizon text workflows; Mirror CLI exposes tool approvals and resumable work. The release says self-distillation recipes will follow in the coming days, so that recipe publication is a promise at this capture. (source)

Compared To

Muse Spark Open is another recent open-weight release, but its license and multimodal design differ. This page does not infer a ranking across different evaluations. (source) (Meta source)

Sources

  • Release announcement. (source)
  • Final model card, read 2026-10-18. (source)

Referenced by

Sources