Beam questions, answered

Beam is a large language model from Reflection AI, announced on October 5, 2026. It is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active per token, built for coding, reasoning and agentic work. It is Reflection's first open-weight model.

Reflection AI, a New York lab founded in 2024 by Misha Laskin (CEO) and Ioannis Antonoglou, both former Google DeepMind researchers. Its investors include Nvidia, Sequoia and Citigroup.

Beam is open-weight. Reflection says it will release the weights under the Apache 2.0 license, which allows commercial use, modification and redistribution. The training data is not open: part of it comes from proprietary licensed datasets.

Reflection says “later this month”, meaning October 2026, together with a technical report, model card and tools for running, evaluating and fine-tuning. No exact date has been given. Our release tracker shows the current status.

Join the waitlist for Reflection's beta API. Approved users get an OpenAI-compatible endpoint at https://api.reflection.ai/openai/v1 with the model ID Beam-501B-A23B. Once the weights are out you will also be able to run it yourself or through hosting providers.

The weights will be free to download and use under Apache 2.0. Running a 501B model is not free, though: you pay for the GPUs or for a hosting provider. Reflection has not published prices for its API.

No official requirements yet. By our estimate the weights alone take about 1 TB in BF16, about 500 GB in FP8 and about 250 GB at 4-bit, plus memory for the KV cache. That means a multi-GPU server; it will not run on a typical laptop or a single consumer GPU.

Not on most published numbers. Against DeepSeek V4.1 Flash, Beam wins on CritPt and the AA Omniscience index but trails on the other seven shared benchmarks, including Terminal-Bench, DeepSWE, HLE and AutomationBench. Reflection's pitch is efficiency and a US-based open model, not a clean sweep.

No. Reflection's own table shows Kimi K3 ahead on 13 of the 14 benchmarks both models reported, with Beam ahead only on τ³ banking (38.0 vs 37.1). Reflection says Beam's advantage over models like Kimi K3 is inference efficiency.

Software engineering and agentic coding. It scores 80.9 on SWE-Bench Verified, 78.0 on SWE-Bench Multilingual and 80.1 on Terminal-Bench v2.1, and it can use tools, browse and call other services in long multi-step tasks.

No. Beam is text-only. It can work with other kinds of data only when they are converted to text, for example through OCR.

The beta API lists 256K tokens with up to 128K output tokens, and Reflection notes this may change during the beta. Its announcement says midtraining extended the effective context to 1 million tokens.

The Wall Street Journal reported in March 2026 that some investors used the phrase. Like DeepSeek, Reflection focuses on efficient, openly released models; unlike DeepSeek, it is a US company, which matters to buyers who prefer not to depend on Chinese models.

Not yet, as of the date at the top of our release tracker. Reflection plans to launch the open weights with distribution partners and integrations with open-source libraries.

No. ReflectionBeam is an independent guide. We are not affiliated with Reflection AI. We link to official sources for every number we publish.