Reflection-Beam-1.0
A large-scale open-source Mixture-of-Experts (MoE) language model released by Reflection AI, with 501B total parameters and 23B active parameters. It supports a 1M-token context window and integrates the Harmony-format reasoning and tool-calling protocol, with Day-0 inference support from vLLM and training support from SkyRL.
Timeline (last 90 days)
Reflection-Beam-1.0 achieved Day-0 support on vLLM, meaning it can run on vLLM on the day the model is released.
Reflection-Beam-1.0 has a total parameter scale of 501B, with only 23B parameters activated per inference.
Reflection-Beam-1.0 is equipped with a context window of up to 1M tokens.
Reflection-Beam-1.0 uses a mixture-of-experts (MoE) architecture (inferred from the difference between total and activated parameters).
Reflection-Beam-1.0 combines a hybrid attention mechanism to handle ultra-long sequences.
Reflection-Beam-1.0 integrates Harmony-format reasoning and tool-use capabilities.
SkyRL (developed by NovaSkyAI) provided Day-0 reinforcement learning training support for Reflection-Beam-1.0, covering LoRA fine-tuning, BF16 full-parameter fine-tuning and MXFP8-precision RL training, and uses vLLM for rollout sampling.
Key data such as Reflection-Beam-1.0's benchmark performance, actual inference latency and effective recall rate at 1M context have not yet been disclosed.
All Facts (8)
Reflection-Beam-1.0 achieved Day-0 support on vLLM, meaning it can run on vLLM on the day the model is released.
50%UnverifiedReflection-Beam-1.0 has a total parameter scale of 501B, with only 23B parameters activated per inference.
50%UnverifiedReflection-Beam-1.0 is equipped with a context window of up to 1M tokens.
50%UnverifiedReflection-Beam-1.0 uses a mixture-of-experts (MoE) architecture (inferred from the difference between total and activated parameters).
50%UnverifiedReflection-Beam-1.0 combines a hybrid attention mechanism to handle ultra-long sequences.
50%UnverifiedReflection-Beam-1.0 integrates Harmony-format reasoning and tool-use capabilities.
50%UnverifiedSkyRL (developed by NovaSkyAI) provided Day-0 reinforcement learning training support for Reflection-Beam-1.0, covering LoRA fine-tuning, BF16 full-parameter fine-tuning and MXFP8-precision RL training, and uses vLLM for rollout sampling.
50%UnverifiedKey data such as Reflection-Beam-1.0's benchmark performance, actual inference latency and effective recall rate at 1M context have not yet been disclosed.
50%