Reflection AI
(New York, NY)
Open-weight frontier agent lab behind Beam (501.3B total / 23.4B active MoE, Apache 2.0) and the open-source Mirror CLI agentic harness; certified consortium member of the DOE Genesis Mission.
What is the latest AI model from Reflection AI?
Reflection AI's primary flagship model is Beam (501.3B Total / 23.4B Active MoE (256 experts), 1,000,000 tokens context). Reflection's first open-weight model: 501B/23B-active MoE for coding, reasoning, and long-horizon agentic work, Apache 2.0 with BF16/FP8/NVFP4 weights plus a speculative-decoding drafter; token-efficient (~3x fewer reasoning tokens for comparable quality), SWE-bench Verified 81.6%, Terminal-Bench v2.1 80.1%, SWE-Bench Pro v2-Hard 76.5%; ~1.3B RL sandboxes across training; $0.25/$1.00 per 1M tokens. How we define “latest”