Beam
Beam 1.0 is Reflection's 501B / 23B-active MoE. OpenTryOn calls the first-party OpenAI-compatible chat API. The model is text in, text out.
| Registry id | Service | Upstream id | Auth |
|---|---|---|---|
beam | understand | Beam-1.0-501B-A23B | REFLECTION_API_KEY |
Base URL is https://api.reflection.ai/openai/v1. Override with REFLECTION_BASE_URL or --base-url.
Reasoning is always on. --reasoning-effort accepts low, medium, high, xhigh, or max. Omit it to use the default your Reflection org reports.
opentryon understand --model beam \
--prompt "Write a 3-sentence lookbook caption for a linen trench."
opentryon understand --model beam \
--prompt "Tighten this PDP copy." \
--system "You edit fashion product copy. Keep claims factual." \
--reasoning-effort high
Images and video are rejected. Use kimi-k2.6, moondream-3.1, or nemotron-omni for media.
The Apache-2.0 weights are on Hugging Face (reflection/Beam-1.0-501B-A23B). Reflection's serve guide asks for four B300 GPUs and a few hundred gigabytes of NVFP4 weights, so there is no local registry id.
MCP tool: understand_beam. Named-model-only — the understand default stays kimi-k2.6.