NVIDIA Nemotron 3.5 Bolt E4B 4GB
A downloadable conversational model for local assistants, interactive characters, and applications that use tools.
Reflection Beam 1.0
reflection.ai
An open-weight text model for coding, reasoning, and tool-using applications, available through hosted and self-hosted deployment.
Visit website
Pricing
Free
Available on
Hosted API, Self-hosted, Linux
Category
Llm Models
What is Reflection Beam 1.0?
Reflection Beam 1.0 is an open-weight model for software engineering, reasoning, and agentic applications. It accepts and generates text and was developed predominantly for English. Its sparse mixture-of-experts architecture has approximately 501B total parameters and 23B active parameters.
Developers can integrate Reflection's hosted API or serve the downloadable weights on their own infrastructure. The release includes BF16, FP8, and NVFP4 variants under Apache 2.0, with deployment recipes for SGLang, vLLM, and TokenSpeed.
Key features
Large context and output budgets
The API documents a 1M-token context and a maximum output of 128K tokens for Beam-1.0-501B-A23B.
Tool calling
The hosted interface supports tool selection and parallel tool calls for application integrations.
Structured and streamed responses
Applications can request JSON objects or JSON Schema outputs and stream responses.
Adjustable reasoning effort
Choose low, medium, high, xhigh, or max effort. Reasoning remains enabled at every setting.
Self-hosted serving options
Official recipes cover downloading weights, starting a server, and checking the deployment with Beam Verifier.
Who is it for?
Coding assistant integration
Evaluate Beam as the text model behind a software development assistant.
Document-heavy reasoning
Explore reasoning over substantial collections of supplied text.
Structured application workflows
Combine tool calls and structured responses within a developer-controlled application.
How to use Reflection Beam 1.0
- Create a project API key in the Reflection platform.
- Configure an HTTP client or compatible SDK for Reflection's API endpoint.
- Select Beam-1.0-501B-A23B and send a chat-completion request.
- Choose reasoning effort and output settings for the task.
- Check the platform's usage allocation before expanding the integration.
Reflection Beam 1.0 pricing
Free
Reflection's API is free during its public beta, subject to daily usage limits. Downloadable weights use Apache 2.0; self-hosting requires separately funded infrastructure.
Check current pricing ↗What to know before you use it
- The API accepts text messages; image, audio, and file inputs are unsupported.
- Compatibility covers Chat Completions and Models. Clients requiring Responses, Embeddings, Batch, or Assistants need adaptation.
- Self-hosting requires substantial infrastructure. The reference setup specifies four NVLink-connected NVIDIA B300 SXM GPUs on Linux amd64, plus Docker and NVIDIA Container Toolkit.
- Public beta API access has daily usage allocations. Requests are declined after allocation is exhausted, and Reflection may change limits or discontinue the beta.
Frequently asked questions
Can I use Beam without self-hosting?
Yes. Reflection provides a hosted API with a documented quickstart using an HTTP client or compatible SDK.
Can reasoning be switched off?
No. Beam always reasons, although the API lets you adjust reasoning effort.
Do 23B active parameters mean it fits on a small GPU?
The full model contains approximately 501B parameters. Reflection's reference self-hosting setup uses four NVIDIA B300 SXM GPUs.