Reflection Beam: Open Weights, API Access and Running Costs

Reflection’s October 16 release provides open weights and a public-beta API signup. Sign in to its platform to configure access, or download an official checkpoint for your deployment. Mirror CLI is a separate beta coding interface.

The OpenAI-compatible base URL is https://api.reflection.ai/openai/v1 and the documented model ID is Beam-1.0-501B-A23B. Query the Models endpoint for your organization’s availability and defaults. Reflection lists a 1M context window and 128K maximum output; an account or serving route can impose a lower limit.

For repository work, begin with a bounded issue and inspect the resulting diff and tests. The coding harness supplies files and tools around Beam; the model itself accepts and returns text.

Reflection’s announcement shows Beam building a live NYC subway map from public MTA data: it searches documentation, checks authentication requirements, finds route geometry and builds the frontend and backend. Another example connects Beam to OpenCode and asks it to research Unsloth and prepare a Gemma-4 Text2SQL fine-tuning notebook.

The p5.js astronaut game demonstrates code generation for a visual application. Beam accepts text input, including OCR output supplied by tools.

The announcement includes initial requests, resulting outputs, setup and user iterations. Read those alongside the demo to understand the surrounding tools and follow-up work. For a trial of your own, use one representative repository issue and compare the accepted change, failed attempts and total calls with your current model.

Reflection’s launch table reports coding and terminal benchmarks for Beam and named models. Beam is close to GLM 5.2 on Terminal Bench v2.1, while the table’s Kimi K3 and DeepSeek V4.1 Flash scores are higher on DeepSWE v1.1. Results depend on the benchmark and agent setup; use the table to choose a comparison for your own task.

Reflection also reports 3–4× less inference compute than GLM 5.2 on advanced reasoning benchmarks. Its efficiency figure estimates generation forward-pass FLOPs as 2 × active parameters × mean generated tokens. Token prices, reasoning length and retries determine the API bill.

Hosted billing depends on the provider and account. Check the rate offered by Reflection’s platform or your chosen provider before sending a batch; the public model card describes weights and capabilities rather than a universal service tariff.

For an agent task, retain input, reasoning and output usage across every call, including tool results and retries. Apply the route’s token rates to that complete record. For self-hosting, compare GPU rental time or owned-machine electricity and depreciation with the cost of an accepted result.

The October 16 release supplies BF16, FP8 and NVFP4 checkpoints under Apache 2.0. Start from the matching official repository and its supported deployment instructions; a lower-bit checkpoint requires a runtime and GPU backend that implement its numerics.

At 501.3B total parameters, a simple weight-only calculation is about 934 GiB at two bytes per parameter, 467 GiB at one byte and 233 GiB at half a byte. These are arithmetic baselines before scales, metadata, KV cache and runtime allocations. They are not measured footprints or proof that a particular GPU configuration will serve the model.

The 23.4B active figure affects computation per token, while inactive experts still need storage and a memory-placement strategy. Allocate the checkpoint, test a short request, then increase context and concurrent traffic while watching memory and latency. A model’s 1M context specification does not mean every hosted provider or local configuration supports a 1M request.