Reflection Beam: 501B open-weight coding model, Apache 2.0, on AWS

Reflection has released Beam, an open-weight model for agentic coding, and AWS published a guide on 16 October 2026 for running it on Amazon SageMaker. The model card on Hugging Face lists 501.3 billion total parameters, 23.4 billion active per token, 256 experts and a 1 million-token context window. It is licensed under Apache 2.0, and Hugging Face shows the first repository was created on 7 October 2026.

What the model card says

• Variants: BF16, FP8 and NVFP4 checkpoints, all instruction-tuned and text-only, with a knowledge cutoff of 30 June 2026.
• Claimed strengths: Reflection says Beam reaches comparable solution quality to leading open models using roughly 3 times fewer reasoning tokens. It does not say which models, and I could not check it.
• Use: the card describes resolving GitHub issues, CLI and DevOps work, and tool use through MCP (Model Context Protocol).
• Licence: Apache 2.0. The card adds that Reflection does not prohibit distillation, fine-tuning, quantisation or using outputs to train other models.

What Reflection's numbers show, and what they leave out

Reflection ran its own evaluations at temperature 0.7 and top-p 0.99. For BF16 it reports 81.6% on SWE-bench Verified, 66.1% on SWE-bench Pro v1, 80.1% on Terminal Bench v2.1, 80.8% on MCP Atlas and 36.1% on Tau3 Banking. The card lists no competing models, so the table cannot tell you whether Beam beats what you run today.

The more usable finding is the quantisation cost. NVFP4 scores 78.4% on Terminal Bench v2.1 against 80.1% for BF16, and 77.5% against 78.3% on SWE-bench Multilingual. A few results move the other way, such as SWE-bench Pro v2-Hard, where NVFP4 scores 78.6% and BF16 76.5%. That is Reflection's data on its own benchmarks; recheck on your tasks after changing precision, as AWS also advises.

On safety, Reflection's in-house Iris evaluation reports that BF16 refuses or pivots on 88.9% of harmful requests and complies with 81.5% of harmless ones. Put the other way, about 1 in 5 harmless requests was not complied with, in a suite Reflection built and has not yet open-sourced.

What it takes to run on AWS

AWS says Beam's 501 billion parameters require a multi-GPU deployment. For FP8 it names ml.p5.48xlarge (NVIDIA H100) or ml.p5e.48xlarge and ml.p5en.48xlarge (H200); for NVFP4 it names ml.p6-b200.48xlarge (B200). At launch the model is available in US East (N. Virginia) and US East (Ohio). The post gives no hourly price and says to delete resources afterwards to avoid ongoing charges.

My own arithmetic, not AWS's: weights alone come to roughly 1 TB in BF16, 500 GB in FP8 and 250 GB in NVFP4, before cache and runtime overhead. AWS describes two paths: a SageMaker JumpStart real-time endpoint, or HyperPod Inference on Kubernetes using JumpStart, SGLang or vLLM. Its sample deploys a short test endpoint, sends one request and cleans up. For a coding workload it suggests recording p50 and p95 time to first token, and tokens per second.

One detail to catch: AWS's smoke test uses temperature 1.0 and top-p 0.7, while Reflection's benchmarks used 0.7 and 0.99. Pick one and test with it before comparing.

What a team should do

• Run your own repositories through the FP8 or NVFP4 variant before paying for BF16 capacity, since Reflection's table shows small gaps.
• Get GPU quota for the instance type first; AWS lists it as a prerequisite.
• Treat the 3 times fewer reasoning tokens claim as unverified until you measure tokens per solved task against your current model.
• Unknown from these sources: AWS pricing, independent benchmark results, and availability outside the two US East Regions.