Beam

Reflection AI's first open-weight model, a 501B-parameter MoE positioned as a lower-compute Western answer to Chinese open models.

Reflection AI · Open-weight/community models · model

Published in Feed

Overview

Reflection AI unveiled Beam on October 5, 2026, its first frontier open-weight model: a 501-billion-parameter mixture-of-experts design that activates only 23B parameters per token, pretrained on 23.8 trillion tokens in under four weeks on 6,144 Nvidia GB300 GPUs, with a 1-million-token context window. Reflection reported scores of 80.1 on Terminal-Bench v2.1, 80.9 on SWE-bench Verified, 77.2 on SWE-bench Pro v2-Hard, 97.8 on AIME 2026, 36.2 on Humanity's Last Exam (no tools), and 90.5 on GPQA Diamond, placing Beam roughly at GLM-5.2's level on most evaluations and behind Kimi K3, GLM-5.3, and DeepSeek V4.1 Flash on the hardest agentic-coding benchmarks, while using an estimated three to four times less inference compute than GLM-5.2 for comparable reasoning scores. Reflection's site lists a 'Beam is now available' post dated October 16, 2026, and secondary coverage (Unite.AI, 80aj) reports the Apache 2.0 weights went live on Hugging Face that day with Baseten day-0 support; Reflection's launch post still points to a Hugging Face collection and API keys on platform.reflection.ai, and the weight release is not yet confirmed on a primary page.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • 501B-parameter mixture-of-experts model, 23B active parameters per token
  • 1,000,000-token context window
  • Pretrained on 23.8 trillion tokens on 6,144 Nvidia GB300 GPUs in under four weeks
  • 80.9 on SWE-bench Verified, 97.8 on AIME 2026, 90.5 on GPQA Diamond
  • Claimed 3-4x lower inference compute than GLM-5.2 at comparable reasoning scores
  • Apache 2.0 open weights reported live on Hugging Face from 2026-10-16 (secondary sources)

Use cases

  • Agentic coding and terminal-driven software tasks
  • Lower-cost self-hosted inference versus comparable Chinese open models
  • Advanced reasoning and long-context workloads

Pricing

Apache 2.0 open weights reportedly published Oct 16, 2026; API pricing unknown (platform.reflection.ai issues API keys).

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
5 Oct 2026
Record updated
20 Oct 2026
Record last checked
20 Oct 2026

Explore the Feed

Other entries filed under open-weight/community models; inclusion does not imply equivalent features.

  • Muse Spark Open

    Meta's open-weight 816B-parameter MoE model (42B active, 1M context) for coding and long-horizon agent work, released Oct 15, 2026.

  • Nemotron 3.5 Bolt

    NVIDIA's open 74B MoE model (9B active) packed into 24GB for local agents on RTX Spark; released Oct 15, 2026.

  • Jais 3

    Six open Arabic-focused language models from 900M to 375B parameters with 512K context, plus Jais Speech for Arabic recognition.

Browse the Feed