Reflection Beam

Name: Reflection Beam
Brand: Reflection AI

Reflection AI, United States, reflection.ai

A 501B open-weight model that US-based Reflection AI released in October 2026. It drew attention as a US-made alternative to Chinese open models, with low inference compute as its strength

Open Models

Go to official siteRequest a correction

CurrencyUSD KRW EUR

Personal starting priceFreeNo paid plan

Free tierFree

Korean supportLimited

Price checked 2026-10-17, source: reflection.ai/blog/beam-is-available. Prices and benefits may change. Check the official site before you pay.

Pros

A large US-made open-weight model (Apache 2.0)
Strong coding benchmarks (SWE-bench Verified 80.9, official announcement)
GLM-5.2-class reasoning with 3 to 4 times less compute (official announcement)
Fine-tuning and evaluation tools to be released as well

Cons

API price and free limits not disclosed
Text only (no image support, etc.)
No chat app for general users
Lower than Kimi K3 and DeepSeek V4.1 Flash on some benchmarks
No public Korean performance data

These are opinions compiled by Pickx from public sources.

At a glance

Beam is the first open-weight model announced on October 5, 2026 by Reflection AI, an AI company in Brooklyn, USA. It has an MoE structure with 501B total and 23B active parameters, ranks near the top among US-made open models on coding benchmarks, and the company says it delivers GLM-5.2-class reasoning with 3 to 4 times less compute. On October 16, 2026 the weights were released on Hugging Face under Apache 2.0 (BF16 / FP8 / NVFP4, including an auxiliary model for speculative decoding), and its own API is open as a public beta that anyone can sign up for and use.

Full model list

Model	Type	Features	Plans that can use it
Beam	Large language model (MoE, text only)	501B total 23B active SWE-bench Verified 80.9 SWE-Bench Pro 65.5 Terminal Bench v2.1 80.1 SWE-bench Multilingual 78.0 (all benchmarks announced by the company)	Free weights Reflection API (public beta)

Plans

Plan	Price	What you can do	Usage limit	Features
Open weights	Free (Apache 2.0)	Download run on your own servers fine-tune	Official limits not disclosed	Released on Hugging Face 2026-10-16 BF16 FP8 NVFP4 includes speculative decoding auxiliary model
Reflection platform	Not disclosed	Sign up at platform.reflection.ai and use the API	Can start free limits not officially disclosed	Public beta also offered on partner platforms

Who should choose this plan

Open weights: Companies and public institutions that are uneasy about Chinese models and want to run a high-performance US-made open model on their own infrastructure.
Reflection platform: Development teams that want to check performance through the API first without running servers.
It is not suitable for general users looking for a conversational app.

Notes

Information checked on: 2026-10-17
The benchmark and efficiency figures are Reflection AI's own announcements (updated October 8). The API price has not been disclosed yet.
Sources: Reflection AI official announcement, Weights release announcement, Hugging Face Beam collection, TechCrunch report

Good for

1 views

Prices converted at 1 USD = ₩1,329.49 (2026-10-19). Original price shown in brackets. Taxes as listed by each provider.

Comments

No comments yet. Be the first.

Your comments and nickname are public. If you see abuse or advertising, report it here and we will review and remove it.