Reflection releases Beam: a 501B parameter open-weight model for agentic coding

What happened

Reflection, a new US-based laboratory, has released Beam, an open-weight mixture-of-experts (MoE) model designed for agentic coding and long-context reasoning [2]. The model features 501 billion total parameters, with 23 billion parameters active per token [2]. It supports a context window of 1 million tokens and accepts text as input [2].

Beam is released under the Apache 2.0 license and is available on Hugging Face [2]. To optimize performance, the model includes a dedicated open-source Rust-based harness [2]. It also features day-0 support for several libraries, including transformers, SGLang, and vLLM [2]. For users running the model on Blackwell hardware, the NVFP4 checkpoint can fit on four B300s [2].

Developers can customize the model using several methods provided by the Reflection team and RadixArk [2]. These include:

• Supervised fine-tuning (SFT) using TRL’s SFTTrainer, which can involve training LoRA adapters with PEFT to manage memory [2].
• Reinforcement learning (RL) using recipes built with Miles, which allows for full-parameter RL or training LoRA adapters [2].
• A specific recipe developed by Reflection and RadixArk that uses GRPO in Miles with BF16 [2].

In benchmark testing, the model achieved the following scores:

• SWE Bench Pro v2-Hard: 76.5% (BF16), 78.4% (FP8), and 78.6% (NVFP4) [2].
• SWE Bench Pro v1: 66.1% (BF16), 65.9% (FP8), and 65.6% (NVFP4) [2].
• Terminal Bench v2.1: 80.1% (BF16), 78.9% (FP8), and 78.2% (NVFP4) [2].
• SWEBench Verified: 81.6% (BF16), 80.2% (FP8), and 80.7% (NVFP4) [2].
• DeepSWE v1.1: 43.3% (BF16), 43.0% (FP8), and 42.7% (NVFP4) [2].
• SWE Atlas Codebase QnA: 31.8% (BF16), 32.2% (FP8), and 31.4% (NVFP4) [2].
• SWEBench Multilingual: 78.3% (BF16), 78.1% (FP8), and 77.5% (NVFP4) [2].

Who is affected

This model is intended for teams that require a frontier-scale model on their own infrastructure, cloud, or hardware [2]. It is specifically built for agentic coding use cases and long-context reasoning [2].