Reflection AI

Reflection is building a full open AI stack for developers, enterprises and public-sector deployments, with collaborations including Dell AI factories, the U.S. Department of Energy and Shinsegae sovereign cloud.

Reflection is building a full open AI stack for developers, enterprises and public-sector deployments, with collaborations including Dell AI factories, the U.S. Department of Energy and Shinsegae sovereign cloud.

Reflection's October 5, 2026 Beam preview documents a sparse mixture-of-experts architecture with 501 billion total and 23 billion active parameters. It describes multi-teacher on-policy distillation combining a capability teacher and separate safety/alignment teacher. Weights remain planned for later October, not released as of observation.

Reflection reports building a frontier-scale LLM and reinforcement-learning training stack and a team with experience from PaLM, Gemini, AlphaGo, AlphaCode, AlphaProof, ChatGPT and Character AI.

Reflection reports building a frontier-scale LLM and reinforcement-learning training stack and a team with experience from PaLM, Gemini, AlphaGo, AlphaCode, AlphaProof, ChatGPT and Character AI.

Reflection states that its open-model stack is validated on Dell AI Factory with NVIDIA for on-premises production and proprietary-data customization.

Reflection raised $2B and is building an open-model stack, production AI infrastructure, enterprise solutions and sovereign/public-sector deployment capabilities.

Reflection raised $2B and is building an open-model stack, production AI infrastructure, enterprise solutions and sovereign/public-sector deployment capabilities.

Reflection's Beam article documents multi-stage data curation, structured code repositories and long-horizon tasks for one-million-token midtraining, alongside adversarially iterated safety datasets and SFT/RL pipelines. These concrete data-engineering and evaluation methods support portable frontier-training experience.

Reflection work spans foundation-model training, reinforcement learning, open model infrastructure, enterprise deployment and public-sector AI stacks, creating broadly portable frontier systems experience.

Reflection work spans foundation-model training, reinforcement learning, open model infrastructure, enterprise deployment and public-sector AI stacks, creating broadly portable frontier systems experience.