Reflection releases Beam, a 501B open-weight MoE model with 1M-token context
In short: Reflection, a new US AI lab, has launched Beam, its first model, as an open-weight mixture-of-experts system with 501B total parameters (23B active per token) and a 1M-token context window, released under Apache 2.0. The model is built for coding and agentic use cases, ships with a custom Rust-based inference harness, and has day-0 support in transformers, vLLM and SGLang. Reflection reports strong results on coding, reasoning and tool-calling benchmarks across BF16, FP8 and NVFP4 checkpoints.
This summary was generated automatically by AI from Hugging Face's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1First model release from Reflection, a new US AI lab
- 2501B total parameters, 23B active per token (MoE architecture, 256 experts, 8 routed per token)
- 31M-token context window using hybrid sliding-window/global attention
- 4Apache 2.0 license, with BF16, FP8 and NVFP4 checkpoint formats
- 5Native tool calling with a dedicated grammar and response template, five reasoning-effort levels
- 6Day-0 support in transformers, vLLM and SGLang, plus DFlash2 speculative decoding
| Parameter | Before | Now |
|---|---|---|
| Total parameters | Not specified | 501B (MoE) |
| Active parameters per token | Not specified | 23B |
| Context window | Not specified | 1,000,000 tokens |
| License | Not specified | Apache 2.0 |
| Tool calling | Not specified | Supported via dedicated grammar and response template |
| Reasoning effort levels | Not specified | low, medium (default), high, xhigh, max |
| Checkpoint formats | Not specified | BF16, FP8, NVFP4 |
Why it matters
It is a rare fully open, frontier-scale model that teams can self-host and fine-tune, giving developers an alternative to closed APIs for agentic and coding tasks, though it demands serious GPU infrastructure (multiple B200/H200-class cards).
What it means for AI agents and contact centers
Beam is text-only and extremely hardware-intensive, so it is not a direct fit for real-time voice agents, but its agentic tool-calling design and long context could be worth tracking for backend reasoning or complex workflow orchestration tasks if self-hosting capacity exists.
Sources
- Hugging FaceOfficialPrimary sourceOriginal article →„Welcome Beam: a 501B open model you can own, customize, and scale“16 Oct 2026, 03:00
- AWS Machine LearningOfficialOriginal article →„Introducing Reflection Beam: an open-weight model for agentic coding on AWS“16 Oct 2026, 18:23
- Published by source
- 16 Oct 2026, 03:00
- Found by our system
- 16 Oct 2026, 16:11
- Summary generated
- 16 Oct 2026, 16:13
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy