跳到正文
Baseten 工程博客·· 2026-10-16AI 評分53

Beam 為編碼與智能體工作負載提供高效推理

Beam: Efficient reasoning for coding and agentic workloads

AI 導讀

Reflection 推出 Beam,這款自回歸 Transformer 模型面向編碼及智能體工作負載。Beam 採用 MoE 架構,總參數量為 501B、每個 token 啟用 23B 參數,上下文窗口最高達 1 million tokens。

正文

Beam is an autoregressive transformer model that accepts text inputs and generates text outputs (mainly in English). It has a mixture-of-experts (MoE) architecture with 501B total parameters, 23B active parameters per token, and a context window of up to 1 million tokens.

Beam is the first in a new family of models from Reflection, with multilingual and multimodal capabilities planned for future releases. Beam is designed for developers and researchers building coding and agentic systems, as well as enterprises that want to customize and run their own models. It works out of the box with the runtimes, frameworks, coding harnesses, and tools developers already use.

Efficient reasoning for coding and agentic workloads

Reflection’s evaluations show strong performance across coding, agentic, reasoning, and STEM benchmarks:

✕

Agentic coding performanceAgentic coding performance

✕

Reasoning performanceReasoning performance

Source: https://reflection.ai/blog/introducing-beam

Reflection trained Beam to solve coding and agentic tasks and use fewer tokens to reach a successful result. The goal is to reduce the cost of these workloads. Reinforcement learning (RL) was a major part of Beam’s training. 

Reflection used longer RL rollouts so Beam could practice solving tasks over multiple steps, using tools and adjusting its approach based on feedback. Reflection rewarded successful solutions during training and applied a length penalty to discourage unnecessary tokens. In short, longer rollouts let Beam work through complex tasks, while the length penalty encouraged it to do so with fewer tokens.

Users can also adjust Beam’s reasoning effort parameter: lower settings favor shorter responses, and higher settings allow more reasoning to improve performance on demanding tasks.

Start using Beam on Baseten

You can use Beam today through Baseten Model APIs and Dedicated Inference. Build coding and agentic systems that solve complex tasks without wasting tokens.

來源:Baseten 工程博客 · baseten.co