Beam (Beam-501B-A23B)
Announced Oct 5, 2026; Apache-2.0 weights and tech report promised for later in October 2026 (no Hugging Face repo yet as of Oct 5). API is beta, reasoning always on (default effort medium), tool calling and structured outputs. No published price. Benchmarks are Reflection's own: SWE-bench Verified 80.9, Terminal Bench v2.1 80.1, GPQA Diamond 90.5, HLE no tools 36.2.
- Context window
- 256,000 tokens
- Max output
- 128,000 tokens
- Knowledge cutoff
- 2026-06
- Input
- text
- Output
- text
- License
- apache-2.0
- Verified
- 2026-10-05
How to call it
| Provider | Model id | Endpoint / URL | Docs |
|---|---|---|---|
| Reflection API (beta, early access / waitlist) | Beam-501B-A23B | https://api.reflection.ai/v1 | docs |
| Reflection API, OpenAI-compatible (Chat Completions, Models) | Beam-501B-A23B | https://api.reflection.ai/openai/v1 | docs |
| Web app / early-access signup | — | platform.reflection.ai/ | — |
Notable capabilities (3)
- Efficient 500B-class open MoE: 501B total / 23B active sparse MoE trained on 23.8T tokens; Reflection says it matches GLM-5.2 on reasoning with 3-4x less inference compute. source
- Large-scale agentic RL: RL on 10.5K GB300 GPUs for 4 weeks: >100M rollouts across ~1M coding, agentic and STEM environments using ~1.3B sandboxes. source
- 1M-token context (training): Pretrained at 256K and extended to 1M tokens in midtraining; the beta API currently exposes 256K context and 128K output. source
Reflection AI's first model. Until the weights ship, access is through the waitlisted beta API. The OpenAI-compatible endpoint
works with the official OpenAI SDKs once you change base_url to https://api.reflection.ai/openai/v1 and use a Reflection API key.
When the weights are released, add the Hugging Face repo and switch status to current.
Timeline entry
- Reflection AI unveils Beam, a 501B-parameter (23B active) Apache-2.0 open-weight MoE; weights due later in October ★★★★
On Oct 5, 2026 Reflection AI announced Beam, its first model: a sparse Mixture-of-Experts with 501B total and 23B active parameters, trained on 23.8T tokens, for coding, reasoning and agentic work. It will be released under Apache 2.0 with weights and a tech report "later in October"; for now it…