Reflection Beam 1.0 API: the open-weight 501B MoE on FastMetal

Reflection AI's Beam 1.0 is available on FastMetal under the model id beam-1.0, in chat, through the API and over MCP. Reflection released the weights on October 16, 2026 under Apache 2.0. It is the company's first open-weight model, built for coding, reasoning and long multi-step agent work.

We added it to chat first. On October 21 we moved it to a different host, which opened it to the API and MCP as well. The model id did not change.

What Beam 1.0 is good at

• Coding and agents. Reflection trained it on resolving real GitHub issues, working in a terminal, and long chains of tool calls. Reflection reports 80.9 on SWE-bench Verified.
• Fewer reasoning tokens. Reflection says Beam reaches answers of similar quality to the leading open models while spending about a third of the thinking tokens. Thinking tokens bill as output on a reasoning model, so this shows up directly on your bill.
• A 501B mixture of experts. About 23B parameters are active per token. The weights are on Hugging Face.
• A 1M-token context. We serve it with 1,048,576 tokens of context. It is text only and does not read images.

Before launch we checked tool calling and schema-constrained JSON output on real requests. The model card says it was developed mainly for English; it also answered our Japanese questions in Japanese.

Pricing

Beam 1.0 is billed per million input tokens and per million output tokens, with cached input (such as a repeated conversation history) at a lower rate. Because it reasons before it answers, the thinking counts as output tokens. Current rates are on the model page.

Billing is prepaid: you load a balance by card, from $5, and requests stop when it runs out. See pricing for the details.

Calling it

The API is OpenAI-compatible. Pass beam-1.0 as the model:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "beam-1.0",
    "messages": [{"role": "user", "content": "Here is a stack trace. What is the cause and how do I fix it? ..."}]
  }'

To try it without code, open chat and pick Beam 1.0 from the model list.

model: "auto" does not pick Beam 1.0 for now. It serves only requests that name it.

Compare real answers first

The model page shows Beam 1.0's actual answers to our 43 shared prompts: summaries, bug fixes, translation and more. Put them beside GLM 5.3, which Reflection used as a comparison point, or beside the model you use today. Everything else we serve is in the model catalog.