Reflection Releases Beam Weights: 501B Total, 23B Active, Apache 2.0, With Its Own Table Showing Kimi K3 Still Ahead on Coding

On October 16 Reflection AI published the weights for Beam on Hugging Face under the Apache 2.0 license. Beam is the company's first open-weight model: a sparse mixture-of-experts model with 501B total parameters and 23B active per token, text-only, with a 1M-token context, aimed at coding, reasoning and long-horizon agent work. Weights ship in BF16, FP8 and NVFP4, plus a drafter model for speculative decoding. The model card lists 81.6% on SWE-bench Verified, 80.1% on Terminal-Bench v2.1 and 90.6% on GPQA Diamond (BF16, all self-reported). The comparison table in Reflection's October 5 launch post shows Kimi K3 still ahead on DeepSWE, Terminal Bench and other rows; Beam's pitch is using less compute at inference time. More than 40 partners support it at launch, and Reflection's own API is in a public beta that is free to start.

The weights arrive eleven days later

When Reflection AI unveiled Beam on October 5, the model was available only through a waitlist, with weights promised "later this month." On October 16 the company announced that the weights are on Hugging Face (repository reflection/Beam-1.0-501B-A23B) under Apache 2.0, which allows commercial use and modification. The release includes BF16, FP8 and NVFP4 precisions and a drafter model for speculative decoding.

The numbers on the model card

Beam is a mixture-of-experts model with 501B total parameters and 23B active per token (501.3B / 23.4B on the model card). It is text-only and mainly English, with a 1M-token context and a knowledge cutoff of June 30, 2026. The model card's BF16 results: 81.6% on SWE-bench Verified, 80.1% on Terminal-Bench v2.1, 90.6% on GPQA Diamond and 80.8% on MCP Atlas, all measured by Reflection.

A rough calculation from the parameter count puts the BF16 weights at about 1TB, so running it on a single local machine is not realistic; most deployments will go through the cloud or multi-GPU servers.

The gaps it lists itself

Reflection does not hide the comparison: in the table from its October 5 post, Kimi K3 leads on DeepSWE (68.0 vs 44.4), Terminal Bench v2.1 (88.3 vs 80.1) and HLE without tools (46.9 vs 36.2). Beam's selling point is efficiency. The company says that on advanced reasoning benchmarks it reaches scores comparable to GLM 5.2 with roughly a third to a quarter of the inference compute, estimated as "2 × active parameters × mean generated tokens," excluding prefill and serving overhead.

Who can use it now

More than 40 partners are involved at launch: Transformers, vLLM, SGLang, Unsloth and other frameworks support it from day one; AWS, Google Cloud, Microsoft Foundry, Together AI, OpenRouter and other platforms host it; and Dell and HPE offer on-premises and air-gapped deployments. Reflection's own API is in a public beta that is free to start, with final pricing not yet announced. It adds another open-weight large model from a US company, after Meta's Muse Spark Open, which this site covered earlier.

via: Reflection blog: Beam is now available, Reflection launch post (October 5), Hugging Face model card