Reflection Beam: open models for coding and AI agents

What does Reflection do, in plain English?

Reflection develops AI models and the software needed to build applications around them. Its Beam model takes text, reasons about a task and returns text or code. Developers can call it from an application, connect it to a coding agent, or download the weights and operate the model on their own infrastructure. For example, a team can ask an agent to investigate a failing test, propose a repair and check the result.

Beam is a sparse mixture-of-experts model with roughly 501 billion total parameters and 23 billion active parameters per token. The practical choice is how to access that capability: a hosted API avoids operating the model yourself; self-hosting gives your infrastructure team responsibility for serving it.

Build applications with the Reflection API

The API quickstart starts with a Reflection platform account and a project API key. The full key is displayed once, so store it securely as REFLECTION_API_KEY.

1. Use the OpenAI-compatible base URL https://api.reflection.ai/openai/v1 in your HTTP client or OpenAI SDK.
2. Select the model Beam-1.0-501B-A23B and send a Chat Completions request containing your text instructions.
3. Read the answer from choices[0].message.content. The platform Playground also lets you try prompts and generate request code.
4. Test the response against your application's expected result before adding longer histories or tool access.

Streaming can show the answer as it arrives. Tool calling lets the model request functions your application defines; your application executes them and returns the results. Structured outputs support JSON and JSON schemas. Reasoning effort is adjustable, but Beam always reasons. Input, reasoning and final output all matter when planning token budgets.

Compatibility covers Chat Completions and Models. Check the compatibility guide before migrating an application that depends on Responses, Embeddings, Images, Audio or other endpoints.

Use Mirror CLI for a real coding task

Mirror is Reflection's terminal coding agent. It can read a repository, edit files and run commands using Beam. Install uv first, then install Mirror with uv tool install reflection-mirror. Set your Reflection key, enter a small project directory and start mirror.

A useful first instruction is to run the project's tests, fix one failing test and run them again. Review requested tool actions, inspect git diff, and rerun the tests yourself. A successful agent message alone does not establish that a code change is correct.

The Mirror setup guide supports macOS 15 or later on Apple Silicon and Linux with glibc 2.28 or later on x86-64 or ARM64. Native Windows is unsupported. Pi, OpenCode and Hermes can also connect to Beam through their compatible provider settings.

Download weights and plan self-hosting

Beam's weights were released on October 16, 2026 under Apache 2.0, in BF16, FP8 and NVFP4 formats. The official reference serving guides use four NVLink-connected NVIDIA B300 SXM GPUs on Linux amd64, Docker and NVIDIA Container Toolkit. They download the main checkpoint and a DFlash2 draft model before serving with SGLang, vLLM or TokenSpeed.

Read the weight download requirements before reserving hardware. This is a reference configuration, not a universal minimum for every possible quantization. Use Beam Verifier to check a deployed endpoint. Customization documentation lists LoRA, supervised fine-tuning and reinforcement-learning stacks; detailed recipes for each stack are still forthcoming.

Costs, limits and data decisions

Checked October 20, 2026: Reflection advertises a free-to-start public-beta API. A universal public first-party per-token tariff was not found. Review your account's billing and allowances; hosting partners set their own terms. Free weights still require compute, memory, storage and operational work.

API rate limits are shared across an organization, including its keys and projects. Requests, tokens and concurrency can be limited; reasoning counts toward usage. Daily allowances reset at 00:00 UTC. Follow the rate-limit guidance when handling a 429 response.

The privacy policy says hosted inputs and outputs are collected and may, depending on your choices, be used to improve models. Enterprise content is governed by customer agreements. Check the applicable terms before sending private code or personal data. Beam was developed predominantly for English; evaluate your own language and domain tasks.

Frequently asked questions

Is Beam free?

Weights are available under Apache 2.0, and the first-party beta is free to start. Hardware, hosting and future account charges are separate; this does not establish unlimited free API usage.

Can I use the weights commercially?

Apache 2.0 permits commercial use subject to its license conditions. Read the model license and the terms of any hosted service you choose.

Can it read images or audio?

The current model accepts text and produces text. Convert other inputs into text using a separate, suitable workflow before submitting them.

Can I run it on my own hardware?

Yes, with suitable infrastructure. Follow the published serving requirements and verify the deployment. Installing Mirror on a laptop uses a model endpoint and does not itself run the full Beam checkpoint locally.

Official resources and developer community

• Developer documentation: API, deployment and customization guides.
• Mirror on GitHub: source code and issue reporting.
• Developer Discord: community discussion.
• Reflection on X: official announcements.
• Contact Reflection: enterprise and partnership inquiries.