Reflection AI releases open-weight Beam coding model, AWS adds hosting
A 501-billion-parameter MoE lands on SageMaker, Apache 2.0 licence attached
The facts
- Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active parameters.
- Reflection reports an effective context length of 1 million tokens.
- Weights are on Hugging Face under Apache 2.0, in BF16, FP8 and NVFP4 checkpoint variants.
- Available at launch via SageMaker JumpStart in US East (N. Virginia) and US East (Ohio); FP8 needs H100/H200 instances, NVFP4 needs B200 instances.
Reflection AI has released Beam, an open-weight mixture-of-experts model built for agentic coding. It has 501 billion total parameters and 23 billion active per token, according to the AWS Machine Learning blog.
Reflection reports an effective context length of 1 million tokens. The company says the model is designed for long-horizon tasks: maintaining code, calling tools, interpreting results and completing multi-step workflows. The weights are published on Hugging Face under the Apache 2.0 licence, in BF16, FP8 and NVFP4 checkpoint variants. That gives developers a choice of memory footprint and hardware compatibility.
AWS has made Beam available through Amazon SageMaker JumpStart in two regions, US East (N. Virginia) and US East (Ohio). A managed real-time endpoint is the fastest path to a working HTTPS API. Developers who need Kubernetes-native control over distributed GPU infrastructure can instead run Beam on SageMaker HyperPod Inference, using the AWS-managed JumpStart container or their own SGLang or vLLM runtime.
AWS specifies hardware by checkpoint. FP8 deployments need ml.p5.48xlarge instances with H100 GPUs, or ml.p5e.48xlarge and ml.p5en.48xlarge with H200 GPUs. NVFP4 requires ml.p6-b200.48xlarge instances built on Nvidia B200 chips. BF16 carries the largest weight-memory footprint and needs a multi-node configuration to hold both the weights and runtime overhead.
AWS also details a HyperPod Inference Gateway. It offers an OpenAI-compatible interface, routes requests across replicas, and supports what AWS calls managed tiered KV caching and intelligent routing, among SageMaker Inference capabilities it says were introduced in 2026. The post walks through sample repositories on GitHub for both the JumpStart and HyperPod paths, including scripts to validate and deploy configurations for JumpStart, SGLang, or vLLM runtimes. It instructs users to mirror checkpoints to Amazon S3 or FSx for production caching rather than pulling repeatedly from Hugging Face.
Why it matters
An Apache 2.0, million-token-context MoE at this scale gives teams a self-hostable alternative to closed agentic-coding APIs, provided they can secure multi-node H100, H200 or B200 capacity. The real competition here is for GPU quota, not model quality — AWS's own deployment guide spends more space on instance sizing than on Beam's capabilities.
Primary source
AWS Machine Learning — aws.amazon.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.