# Reflection Beam, Explained

> What Reflection AI's Beam open-weight model is: 501B total parameters with 23B active per token, text-only MoE for coding, reasoning, and agents. Benchmark scores against GLM 5.2/5.3, Kimi K3, Qwen 3.8 Max, and DeepSeek V4.1 Flash, with an interactive benchmark explorer. No published price; Apache 2.0 weights promised later in October 2026. Independent explainer, not affiliated with Reflection AI. Full page: https://lab.awaara.cn/reflection-beam/

## What Beam is

- Beam is the first open-weight model from Reflection AI (New York, founded 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou). Announced October 5, 2026.
- Text-only sparse mixture-of-experts: 501B total parameters, 23B active per token (4.6% activation). Pretrained on 23.8T tokens: 6,144 Nvidia GB300 GPUs in under four weeks, then RL on 10,500 GB300s with over 100M rollouts.
- Context: 1M tokens announced; beta API allows 256K input+output combined, up to 128K output.
- API beta on a waitlist. Base URL https://api.reflection.ai/openai/v1, model id Beam-501B-A23B, OpenAI-compatible for chat completions and model listing only. No weights, no price, no public checkpoint as of October 11, 2026.
- Apache 2.0 weights plus technical report and model card promised later in October 2026.
- A reasoning-effort parameter trades response length for performance. A successor model is already in training.

## Benchmarks (Reflection's own numbers, not independently verified)

- SWE-Bench Verified: Beam 80.9 (only model reported)
- SWE-Bench Multilingual: Beam 78.0 (only model reported)
- SWE-Bench Pro v1: Beam 65.5, Qwen 3.8 Max 67.7, GLM 5.2 62.1
- SWE-Bench Pro v2-Hard: Beam 77.2, Kimi K3 88.2, GLM 5.3 84.3
- Terminal-Bench v2.1: Beam 80.1, DeepSeek V4.1 Flash 90.6, Kimi K3 88.3, GLM 5.3 88.2, Qwen 3.8 Max 86.6, GLM 5.2 81.0
- DeepSWE v1.1: Beam 44.4, DeepSeek V4.1 Flash 74.2, Kimi K3 68.0, GLM 5.3 61.0, Qwen 3.8 Max 51.0, GLM 5.2 44.0
- SWE Atlas Codebase QnA: Beam 34.6, Kimi K3 68.0, GLM 5.3 61.0
- AIME 2026: Beam 97.8. GPQA Diamond: Beam 90.5 (both Beam only, Reflection-reported).
- Read plainly: Beam leads on SWE-Bench Verified and Multilingual, sits level with GLM-5.2, trails the newest open models on most shared rows. The pitch is compute efficiency, not the leaderboard.

## Pricing

- Beam: no published price. No per-token rate card as of October 6, 2026.
- What the rival field charges per 1M tokens, input/output (OrcaRouter, October 6, 2026): DeepSeek V4.1 Flash $0.15/$0.60, GLM 5.3 $1.26/$3.96, GLM 5.2 $1.40/$4.40, Qwen 3.8 Max $2.00/$6.00, Kimi K3 $3.00/$15.00.
- Claims that Beam is cheaper today rest on compute estimates; there is no rate card.

## Status check

- October 4, 2026: Axios reports a Reflection open model is close.
- October 5, 2026: Beam announced.
- October 10, 2026: Financial Times reports Nvidia in early talks to acquire Reflection or invest further (unconfirmed).
- Later October 2026: weights, technical report, model card promised, with distribution partners.

## Links

- Explainer: https://lab.awaara.cn/reflection-beam/
- Benchmark explorer: https://lab.awaara.cn/reflection-beam/#bench
- Availability checker: https://lab.awaara.cn/reflection-beam/#use
- Related: Haiku 5.5 pricing https://lab.awaara.cn/claude-haiku-5-5/ · Jev, explained https://lab.awaara.cn/jev/ · Decisions API https://lab.awaara.cn/decisions-api/