TL;DR
On October 5, 2026, US startup Reflection AI unveiled Beam, its first open-weight model: a MoE architecture with 501 billion total and 23 billion active parameters, built for coding, reasoning and agent tasks. The company says inference compute is one-fourth to one-third that of comparable models (company's own claim).
Event overview
On October 5, 2026, US startup Reflection AI unveiled Beam, its first open-weight model. Per the company's official release, IT Home and InfoQ, Beam uses a MoE architecture with 501 billion total and 23 billion active parameters, targeting coding, reasoning and agent tasks; weights are expected under an Apache 2.0 license later this month.
Key facts
- Event: Reflection AI launches Beam (October 5, 2026, official channels).
- Company background: Reflection AI was founded in March 2024 in New York by former Google DeepMind researchers Misha Laskin (CEO) and Ioannis Antonoglou (President and CTO), and is backed by Nvidia (per IT Home and InfoQ).
- Model specs: MoE with 501B total and 23B active parameters; text-only with 1M effective context tokens; trained end-to-end from scratch (per the company's release).
- Performance claim: Reflection says Beam matches GLM-5.2-level reasoning on coding and agent tasks while requiring one-fourth to one-third the inference compute of comparable models (company's own claim).
- Release plan: weights, technical report and evaluations later this month; early-access waitlist now open (per the company's release).
Official sources
- Reflection AI official release (October 5, 2026): Introducing Beam: Reflection's 501B open-weight model
- IT Home (October 6, 2026): Nvidia-backed Reflection AI unveils its first open-weight model Beam