Reflection AI
US open-weight lab focused on reasoning and agentic models; released the 501B-parameter Beam under Apache 2.0 in October 2026.
Reflection AI was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. It initially focused on autonomous coding agents before repositioning in 2025 as an American open-weight frontier lab, a counterweight to Chinese open models.
Beam was previewed on Oct 5 2026 and its weights were released on Oct 16 under Apache 2.0 in BF16, FP8 and NVFP4. It is a 501B sparse MoE with 23B active parameters, text-only, with a 1M-token effective context (256K was the length used in reinforcement learning) and 23.8T pretraining tokens. The preview table gave 80.9 on SWE-bench Verified; the final model card reports 81.6 (BF16) and 80.1 on Terminal-Bench v2.1, where several rivals score higher. Reflection says Beam matches GLM-5.2 on advanced reasoning with 3-4x less inference compute, an estimate based on theoretical FLOPs from active parameters and generated tokens that leaves out prefill and hosting costs. The scores are Reflection’s own and, per press coverage, were not independently verified.
Cyber
On Nov 17 2026 Reflection launched a Trusted Access Program for defensive cybersecurity, starting with Laser, an autonomous penetration-testing harness for its Mirror terminal agent, offered as a research preview only to accepted commercial partners. Laser scopes each test, records findings in a shared evidence ledger and uses a separate supervisor to review evidence and direct the executor. The gated approach mirrors OpenAI’s Daybreak and Google’s Fairwind Program for Gemini.
Timeline
Every dated event in the wiki that involves Reflection AI, newest first.
- Reflection AI gates its Laser autonomous pentest agent behind a Trusted Access Program
Reflection's program gives vetted organizations early access to selected capabilities for defensive cybersecurity, starting with Laser, an autonomous penetration-testing harness for its Mirror terminal agent offered as a research preview only to accepted commercial partners. Laser scopes tests, keeps a shared evidence ledger and uses a separate supervisor to review evidence and direct the executor. It follows gated cyber releases such as OpenAI's Daybreak and Google's Fairwind Program for Gemini.
- Reflection AI releases Beam open weights
501B sparse MoE under Apache 2.0; 80.1% on Terminal-Bench 2.1 and 81.6% on SWE-bench Verified per the model card, with a claim of 3-4x less inference compute than GLM-5.2 (self-reported).
Sources: BenchmarkList, “Beam 1.0 Benchmark Scores & Evals”; VentureBeat
Connections
Reflection AI and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
What links here
Model family GLM
Person Misha Laskin
Concept Open-weight vs. closed models · Token
Publication VentureBeat
Cited by
Cited as a source by 1 timeline events.
Sources
- BenchmarkList, “Beam 1.0 Benchmark Scores & Evals”, Oct 16 2026
- Layer3 Labs, “Reflection AI Beam Explained: The 501B Open-Weight Model”, Oct 16 2026
- Reflection AI, “Launching Reflection's Trusted Access Program, starting with cyber”, Nov 17 2026