發生什麼事
Built mainly for English, Beam handles up to 1M tokens of context; under the hood it is a sparse MoE that puts only 23B of its 501B parameters to work at any one time (Beam model card). A small draft model for speculative decoding ships alongside the weights, which come in three precisions: NVFP4, FP8 and BF16 (Reflection blog).
Reflection also launched Mirror CLI, a terminal agent tool that Beam was trained alongside; other harnesses it has been tried on include Hermes, OpenCode and Pi. In the cloud, it can be reached through OpenRouter, Ollama, Microsoft Foundry, Google Cloud, AWS and other platforms. Teams that want it inside their own firewall, or even in an environment with no connection to the outside network, get deployment help from HPE and Dell (Reflection blog).
Correction (2026-10-22)
Blog links were added to the bottom line and to the distillation-recipe preview; the platform listed as Microsoft Azure now reads Microsoft Foundry, matching the blog.
01Sources
- Beam is now available: Reflection's official blog, October 16; distillation-recipe preview, free trial, partner platforms, Mirror CLI, precision formats, parameters, weight licensing
- Beam model card: 1M-token context, primarily for English
- Beam 1.0 Hugging Face collection: weights download page
- @reflection_ai launch post: 10/16 13:01 UTC (21:01 Taipei time)