Reflection AI has introduced Beam, a new open-weight artificial-intelligence model built for coding, reasoning and tasks that require an AI agent to use tools. The company says Beam contains 501 billion total parameters while activating 23 billion for each token, a design intended to offer strong performance without using the entire network for every request.
That sounds like a full public release, but there is an important limitation: Beam is still undergoing final safety testing and evaluations. Developers can join an early-access waitlist, while Reflection says the model weights, technical report, model card and developer tools will arrive later in October 2026.
What Reflection announced
Reflection unveiled Beam on October 5 as its first open-weight model. It is a sparse mixture-of-experts, or MoE, system. Instead of running all 501 billion parameters for every piece of text, the model routes each task through a smaller selection representing about 23 billion active parameters.
The basic idea is similar to having a large team of specialists while calling only the relevant experts for a particular job. A sparse architecture can reduce the amount of computation needed during inference, although the real hardware requirement and operating cost also depend on memory, quantization, software optimization, context length and deployment scale.
According to Reflection’s official launch post, Beam was pretrained on 23.8 trillion tokens from web data and proprietary licensed datasets. The company says reinforcement-learning training used 10,500 Nvidia GB300 GPUs for four weeks and generated more than 100 million training rollouts.
Those are manufacturer claims, not measurements independently reproduced by Gadget N Widget. The full technical report and model card were not available at publication time.
Beam is focused on coding and AI agents
Reflection is positioning Beam as a workhorse for software engineering, command-line work, tool use, search and multi-step reasoning. It is a text-only model, but the company says it can work with other types of information after that data is converted into text.
The published benchmark table shows a mixed picture rather than a universal lead. Reflection reports an 80.9 score on SWE-bench Verified, 80.1 on Terminal-Bench 2.1 and 78.7 on MCP Atlas. On some tests Beam is competitive with or ahead of selected open-weight rivals; on others, newer models such as Kimi K3, GLM 5.3, Qwen 3.8 Max or DeepSeek V4.1 Flash score higher.
Reflection says Beam approaches the coding and agent performance of larger open models while using less inference computation. Reuters reports that the launch is aimed partly at competing with lower-cost Chinese models from companies including Z.ai, Alibaba and Moonshot AI.
What “open-weight” actually means
Open-weight does not automatically mean fully open-source. Model weights are the learned numerical values that control how the system processes input and produces output. Releasing them can let organizations run the model on their own infrastructure, inspect aspects of its behavior and fine-tune it for specialized work.
But the practical rights depend on the license, and a weight release does not necessarily include the complete training data, preprocessing pipeline or training code. Reflection has not yet published Beam’s final weight package or accompanying license, so developers cannot fully assess redistribution rules, commercial-use terms or exact deployment requirements.
That distinction matters for companies considering a local deployment. Beam’s sparse design means fewer parameters are active at once, but an organization may still need substantial accelerator memory and infrastructure to store and serve a 501-billion-parameter model. The 23-billion active figure should not be interpreted as meaning the complete model fits wherever a normal 23-billion-parameter model would fit.
Availability, API access and pricing
Beam is not yet available as an unrestricted public download. Reflection says it is completing red-team tests and evaluations, with the weights and supporting materials planned for later this month. No exact release date was provided.
The Reflection API documentation describes the service as a beta that is opening gradually through a waitlist. It offers a native endpoint and an OpenAI-compatible endpoint, which should make basic migration easier for developers already using OpenAI-style chat-completion tools. The documentation also warns that behavior and limits may change during the beta.
Reflection has not published consumer subscription pricing or final API rates for Beam in the launch material reviewed for this article. There is also no general consumer chat app announcement comparable to ChatGPT, Gemini or Claude. For now, the launch is most relevant to developers, AI infrastructure teams and organizations evaluating models they may eventually host or customize.
Safety claims still need scrutiny
Open-weight releases offer more control, but they can also make safeguards harder to enforce after a model leaves the developer’s servers. Reflection says Beam’s safety training included single-turn and multi-turn conversations, jailbreak attempts and scenarios in which a simulated adversary pressures a tool-using model to take unsafe actions.
The company is still conducting final red-team work, which is one reason the weights have not yet been released. Until the model card and independent evaluations arrive, it is too early to judge how Beam handles harmful requests, software-security tasks, prompt injection, unreliable tool output or sensitive enterprise data.
Who should pay attention
- Software developers: Beam is designed around coding, terminal tasks and tool use, but broad access is not open yet.
- Businesses with private data: A future downloadable model could support controlled, on-premises deployments, subject to the license and hardware requirements.
- AI researchers: The promised weights and technical report may make it easier to study the model, reproduce tests and probe its safety claims.
- Everyday users: There is no immediate consumer product or confirmed price, so Beam is not yet a direct replacement for a mainstream AI assistant.
The practical takeaway
Beam is a significant announcement because it combines a very large total parameter count with a much smaller active network and targets the fast-growing market for coding agents. Reflection’s initial benchmark results suggest the model may be competitive in several areas, but those figures come from the developer and do not show a clean win across every test.
The more important moment will come later this month if Reflection releases the weights, license, model card and technical report as promised. Those materials will determine whether Beam is genuinely practical to deploy, how open the release is, what it costs to run and how its safety and performance hold up under independent testing.
Featured image: official Beam launch artwork from Reflection AI.