Reflection AI has launched Beam, its first open-weight model, designed for coding, reasoning and AI agents. The US startup says Beam performs close to some larger rivals while costing less to run. Those are company claims, and the model's weights are not yet available to download.
What happened?
On October 5, 2026, Reflection AI announced Beam in a long blog post. TechCrunch reported that the announcement followed press reports of an imminent launch. Bloomberg described Reflection as an Nvidia-backed startup founded by two former Google DeepMind researchers.
The launch matters because Western open-weight models have struggled to match the strongest Chinese ones. Semafor reported that Beam could give Western companies and governments another option if they cannot or will not use Chinese models.
What is Beam?
Beam is a text-only language model. Reflection says it is a "sparse Mixture-of-Experts" model with 501 billion total parameters, 23 billion of them active at a time. [Reflection AI]
"Open-weight" means the trained model files are released, so developers can download and run the model themselves. This is not always the same as "open-source," because the training data and training code may stay private. "Mixture-of-Experts" (MoE) means the model contains many specialist sub-networks, and only a few activate for each piece of text. Think of a hospital: one patient does not see every doctor. This keeps each answer cheaper than using all 501 billion parameters.
According to Reflection, Beam was trained on 23.8 trillion tokens and supports a 1-million-token context window, roughly the length of several long books. TechCrunch also reports the 1-million-token window.
Why coding and AI agents?
Reflection says it trained Beam with a particular focus on coding and agentic work. An AI agent is different from a chatbot. A chatbot answers a question. An agent takes a goal, plans steps, uses tools such as a terminal or web search, checks the results and keeps going until the task is done.
Reflection says it used reinforcement learning, where a model improves by attempting tasks and being graded. The company reports more than 100 million attempts, nearly one million coding, agent and science environments, and four weeks on 10,500 Nvidia GB300 chips. These figures are company-reported. Help Net Security summarized the same numbers.
What could developers do with it?
Reflection's own demos give the clearest picture. The company says Beam built a live New York subway map from public transit data. In another demo, plugged into the OpenCode tool, it wrote a notebook to fine-tune a small Gemma model. [Reflection AI] These are curated examples, not independent tests.
Reflection also says Beam has a "reasoning effort" setting. Lower settings give shorter answers, and higher settings think longer on hard tasks. [Reflection AI]
How do the benchmarks look?
Reflection says Beam scores comparably to Z.ai's GLM-5.2 on advanced reasoning while using three to four times less inference compute. [Reflection AI] It also says Beam beats Inkling, a rival US open model, on four coding tests where both report scores. [TechCrunch]
Against the strongest open models, the picture is more mixed. In Reflection's own table, Beam scores 80.1 on Terminal Bench v2.1, a command-line test, against 81.0 for GLM 5.2, 88.3 for Kimi K3 and 90.6 for DeepSeek V4.1 Flash. On DeepSWE v1.1, a software-engineering test, Beam scores 44.4, just above GLM 5.2 at 44.0 but well behind Kimi K3 at 68.0 and DeepSeek V4.1 Flash at 74.2. Reflection says the rival scores came from Artificial Analysis and DataCurve.
Reflection itself says Kimi K3 remains ahead on raw capability and that Beam's advantage is efficiency. Help Net Security's headline said Beam trails top open models on coding tests, pointing to the same numbers.
Beam vs. other open-weight models
No clear winner emerges from the data. On size, TechCrunch reports that GLM 5.2 has about 744 billion total and 40 billion active parameters, so Beam is smaller on both counts. On modality, Inkling is multimodal while Beam is text-only. On origin, Beam is positioned as a Western alternative to Chinese models from Z.ai, DeepSeek and Alibaba's Qwen.
On licensing, Beam's planned license is Apache 2.0. I could not verify licensing terms for the other models, so no comparison is made.
Why this matters for AI agents
Agents work in long loops: plan, write code, run it, read errors, fix, repeat. That is expensive. A model that reaches good answers with fewer tokens could make always-on agents cheaper to run, and that is the bet behind Beam's efficiency pitch.
Open weights add control. Companies can run the model on their own systems, which matters for private code and regulated industries. TechCrunch reports that Reflection is aiming at enterprises and governments that want custom local AI systems.
Availability
Right now, Beam is available only through early access, via a waitlist on Reflection's platform, for a select group of users. Reflection says it will release the weights under an Apache 2.0 license later this month, along with documentation, a technical report and a model card. [Reflection AI] It also says Beam will launch with distribution partners and integrations with open-source tools, though no specific providers were named in the sources I reviewed.
The Apache 2.0 license remains a company plan until the files actually appear. No pricing has been published in the sources I reviewed.
Limitations and open questions
TechCrunch noted that Reflection's performance claims have not been independently verified. The "3–4×" compute figure is also an approximation. Reflection's formula uses active parameters and average tokens generated, and it leaves out serving overhead. [Reflection AI] Real costs will depend on your workload and hardware.
Hardware requirements have not been published. A 501-billion-parameter model is unlikely to run on an ordinary laptop, but that is my inference, not a confirmed figure.
Beam is also still in final red-teaming, and Reflection says safety results will come in its technical report. Help Net Security noted that when given web access, Beam began querying other AI models on its own, behavior worth logging. Finally, Beam is text-only, so it cannot directly process images.
Conclusion
Beam is a notable Western open-weight release, but it is not a clear leader. Its case rests on efficiency, not top scores. Whether that case holds depends on the weights, the technical report and independent testing, all expected within weeks. Developers should wait for those before making decisions.
7. Sources
Official
Independent coverage
Help Net Security: Beam trails top open models on coding tests
Semafor: Reflection AI unveils an open-source Western answer to Chinese labs
