October 6, 2026
Beam connects to coding agents: Reflection launches API access via waitlist
Reflection's first model is available through the API via a waitlist, and the company promises to release the weights in October.
Beam scored 80.9 on SWE-bench Verified in Reflection's Oct 5 evaluations. This benchmark tests how well a model fixes bugs in repositories.
Beam is Reflection's first model with a promise of open weights. At the Oct 5 announcement, the company launched access through a beta API and scheduled the weights for release at the end of the same month.
The model handles text and has 501 billion parameters, with 23 billion active per request. In its Mixture-of-Experts architecture, only part of the model processes each request.
On the terminal benchmark, Beam trailed two models in Reflection's comparison. The company's Oct 5 results for Terminal Bench v2.1:
| Model | Score | | --- | ---: | | Beam | 80.1 | | GLM 5.2 | 81.0 | | Kimi K3 | 88.3 |
How to get access. First, join the waitlist at platform.reflection.ai. Once your account is activated, create a key in API Keys and set it in the `REFLECTION_API_KEY` environment variable.
Beam connects through the OpenAI SDK using a Reflection key. Set `base_url="https://api.reflection.ai/openai/v1"` and `model="Beam-501B-A23B"`.
An agent in the terminal. Mirror CLI runs in the project folder and selects Beam by default. Reflection's Oct 5 instructions give these installation and launch steps:
1. Run `curl -fsSL https://raw.githubusercontent.com/reflection-oss/mirror-beta/main/install.sh | bash`. 2. Open a new terminal. 3. Run `mirror` in your project and enter your API key.
Mirror CLI requires macOS 15 or later on Apple Silicon, or Linux with glibc 2.28 or later on x86-64/ARM64. It does not run natively on Windows.
Reflection also prepared configurations for other agents. The Oct 5 instructions were tested with these versions:
- Pi 1.0.3. - OpenCode 1.18.34. - Hermes v0.21.5.
These beta API configurations specify a context window of 262,144 tokens and a maximum output of 131,072 tokens. OpenCode limits output to 32,000 tokens by default. Longer responses require the environment variable `OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX=131072`.
In Reflection's demo, Beam used OpenCode to prepare an Unsloth notebook for fine-tuning Gemma-4 on Text2SQL, which converts questions into SQL queries. The company reports a 66.5% increase in Gemma's accuracy on a held-out test set.
Original source: [Reflection's Beam announcement, Oct 5, 2026](https://reflection.ai/blog/introducing-beam).
Later in Oct 2026, Reflection promises to release the weights under Apache 2.0, a technical report, a model card, and tools for running, evaluating and fine-tuning the model.