Reflection AI releases Beam, a 501B-parameter open-weight MoE coding model under Apache 2.0

Reflection AI has released Beam 1.0, an open-weight sparse mixture-of-experts (MoE) model with 501 billion total parameters and 23 billion active parameters, aimed at agentic coding and long-horizon software tasks. The weights are published on Hugging Face under the Apache 2.0 license, and AWS has made the model available through Amazon SageMaker JumpStart.

Key specifications

• Architecture: Sparse mixture-of-experts
• Parameters: 501B total, 23B active per token
• Context length: Reflection reports an effective context length of 1 million tokens. This is a company-reported figure, and no independent verification was cited.
• License: Apache 2.0
• Checkpoints: BF16, FP8, and NVFP4, each with a different memory footprint and hardware compatibility
• Per-token API pricing: not yet disclosed
• Benchmark scores: not disclosed in the AWS announcement
• Training cutoff: not disclosed

According to Reflection, Beam is designed for resolving software issues, modifying code, calling tools, and acting on intermediate results. That includes maintaining code and completing multi-step workflows.

Availability on AWS

At launch, Beam is available in SageMaker JumpStart in US East (N. Virginia) and US East (Ohio). Weights can also be pulled from Hugging Face Hub or mirrored to Amazon S3.

AWS describes two deployment paths:

1. SageMaker JumpStart real-time endpoint. This is the managed path, with a packaged model, container, and HTTPS endpoint. An AWS sample script supports all three checkpoint variants. NVFP4 is the default, and FP8 and BF16 are selectable with a flag.
2. SageMaker HyperPod Inference. This is the Kubernetes-native path on an Amazon EKS-orchestrated cluster. It can use the AWS-managed JumpStart container or a custom SGLang or vLLM runtime. HyperPod provides an OpenAI-compatible gateway, model and image caching on node-local NVMe, managed tiered KV caching, and intelligent routing.

Hardware requirements

The 501B total parameter count means every deployment is multi-GPU. AWS lists the following configurations:

Checkpoint	Instances	GPUs
BF16	Multi-node configuration (largest weight-memory requirement)	Not specified
FP8	ml.p5.48xlarge; ml.p5e.48xlarge; ml.p5en.48xlarge	NVIDIA H100 (p5); H200 (p5e, p5en)
NVFP4	ml.p6-b200.48xlarge	NVIDIA B200

AWS advises calculating memory for weights, KV cache, and runtime overhead per replica, then confirming the estimate with a load test. Long-context use will increase KV cache demand, so the 1M-token figure will not be practical on every configuration. Instance pricing is not covered in the announcement. Because the weights are free, the cost of running Beam is the underlying compute.

What this means

Beam is a large, permissively licensed model aimed squarely at agentic coding. Few open-weight models at this scale ship under Apache 2.0. That license allows commercial use, modification, and redistribution without the custom usage restrictions some large open models carry.

The 23B active parameter count matters for serving. Per-token compute should be closer to a mid-sized dense model than to a 501B one. Memory is the constraint, since all 501B parameters must be resident. That explains the focus on FP8 and NVFP4 checkpoints, which cut weight memory and map to H100/H200 and B200 hardware.

The missing piece is evidence. The AWS post contains no benchmark results, such as SWE-bench-style issue resolution scores, and the 1M-token context is only a Reflection-reported "effective" length. Until independent evaluations appear, Beam's coding performance relative to closed frontier models and other open-weight options is unverified.

The launch is also a distribution signal. Day-one JumpStart and HyperPod support gives enterprises a managed route to self-hosting a frontier-scale open model inside their own AWS environment. Availability is limited to two US regions for now.