Reflection
A reflection mechanism in AI agent architectures. After completing a round of actions, the agent compares execution results against expected goals, identifies deviations, and replans accordingly, forming a plan–execute–reflect–replan loop. This mechanism is key to an agent's autonomous continuous improvement.
Timeline (last 90 days)
Reflection Beam has 501B total parameters and 23B activated parameters, uses the Apache 2.0 license with no additional terms, and its smallest official checkpoint is in NVFP4 format at 294GB.
Over the past 72 hours, Meta, Nvidia, Reflection and DeepSeek have successively released open-source large models, with parameter counts ranging from 74B to 816B.
The Reflection team announced on Twitter that it will hold a livestream titled "How to Build Beam in Minutes."
Beam is a product built by the Reflection team.
Reflection announced on October 16, 2025 that Beam is generally available; Beam uses a mixture-of-experts architecture with about 501 billion total parameters and about 23 billion activated per pass, supports a 1-million-token context, is under the Apache 2.0 license, and is offered in BF16, FP8, and NVFP4 versions.
Reflection's safety report rates Beam as Limited across four categories of systemic risk: biological, chemical, loss of control, and cyberattack.
Reflection's concurrently open-sourced Mirror can run Beam in the terminal, inspect agent activity, approve tool calls, and guide tasks; its architecture consists of a Python frontend and a Rust core, and it supports resuming sessions, parallel execution in worktrees, MCP, and Python extensions.
Reflection released Beam, a 501B-parameter open MoE with 23B activated per token, an Apache 2.0 license, a 1 million token context, and a target of coding and agentic tasks.
The SkyRL team announced Day-0 reinforcement learning support for Reflection's Beam model, meaning the Beam model can be plugged into the full RL training pipeline on the day of its release.
Cline's model catalog expanded from 6,952 to 7,041, and two providers, CommandCode and Reflection, were added.
8 more timeline events
All Facts (16)
Reflection Beam has 501B total parameters and 23B activated parameters, uses the Apache 2.0 license with no additional terms, and its smallest official checkpoint is in NVFP4 format at 294GB.
50%UnverifiedBeam is a product built by the Reflection team.
50%UnverifiedReflection's safety report rates Beam as Limited across four categories of systemic risk: biological, chemical, loss of control, and cyberattack.
50%UnverifiedCline's model catalog expanded from 6,952 to 7,041, and two providers, CommandCode and Reflection, were added.
50%UnverifiedThe Agent writes errors into the current conversation's context memory via audit Skills, implementing a lightweight Reflection mechanism that improves task success rates without retraining the model.
50%UnverifiedThe Reflection mechanism causes the agent to compare execution results against expected goals after completing a round of actions, and if discrepancies are found, it replans, forming a cycle of plan—execute—reflect—replan.
50%UnverifiedThe generate–verify–correct automated closed loop typically relies on two types of mechanisms: static checking combined with unit testing, and reflection and self-criticism loops.
50%UnverifiedOver the past 72 hours, Meta, Nvidia, Reflection and DeepSeek have successively released open-source large models, with parameter counts ranging from 74B to 816B.
50%UnverifiedThe Reflection team announced on Twitter that it will hold a livestream titled "How to Build Beam in Minutes."
50%UnverifiedReflection announced on October 16, 2025 that Beam is generally available; Beam uses a mixture-of-experts architecture with about 501 billion total parameters and about 23 billion activated per pass, supports a 1-million-token context, is under the Apache 2.0 license, and is offered in BF16, FP8, and NVFP4 versions.
50%UnverifiedReflection's concurrently open-sourced Mirror can run Beam in the terminal, inspect agent activity, approve tool calls, and guide tasks; its architecture consists of a Python frontend and a Rust core, and it supports resuming sessions, parallel execution in worktrees, MCP, and Python extensions.
50%UnverifiedReflection released Beam, a 501B-parameter open MoE with 23B activated per token, an Apache 2.0 license, a 1 million token context, and a target of coding and agentic tasks.
50%UnverifiedThe SkyRL team announced Day-0 reinforcement learning support for Reflection's Beam model, meaning the Beam model can be plugged into the full RL training pipeline on the day of its release.
50%UnverifiedAmerican company Reflection released an open-source model one day before the launch of Mistral Large 4, and TechCrunch said it aims to challenge Chinese models at a lower compute cost.
50%UnverifiedReflection has launched Beam, a sparse Mixture-of-Experts (MoE) model with a total of 501 billion parameters, activating approximately 23 billion parameters per input.
50%UnverifiedThe Reflection team released an open-weight large language model named Beam, with a parameter scale of 501 billion (501B).
50%