Reflection has moved Beam from preview to a downloadable release. On October 16, the company published its first model’s weights under Apache 2.0, alongside developer tooling and hosted access. Unlike the October 5 preview, builders can now download, customize, and deploy it.
A large model with sparse active computation
Beam is a text-only mixture-of-experts model with 501 billion total parameters, roughly 23 billion active per token, and a one-million-token context window. It targets coding, reasoning, and multi-step tool use.
The smaller active-parameter count describes computation per token—not the full model’s storage requirements. The release includes BF16, FP8, and NVFP4 formats, giving deployment teams choices to evaluate against their infrastructure and workload.
Mirror brings a practical agent workflow
Reflection’s Mirror CLI reads repositories, edits files, and runs commands from the terminal. It supports resumable conversations and separate worktrees for parallel agents, so tasks can proceed in different checkouts.
Mirror remains in early beta; commands and behavior can change. Its tool approvals govern what a developer reviews, but they are not a sandbox. Approved commands, extensions, and local MCP servers run with the user’s account permissions.
That distinction matters alongside execution-infrastructure projects such as Perplexity’s SPACE sandbox platform: a coding agent’s workflow and the environment containing its execution are separate layers.
Why it matters for builders
For teams building apps and websites, including SunMarc App Labs, Beam offers another candidate for real maintenance tasks: fixing bugs, updating pages, or completing small features.
The useful comparison is accepted changes, human corrections, completion time, and total cost—not just benchmark rankings. A bounded repository task with tests and a reviewable diff would be a practical starting point. This is a proposed evaluation, not a claim that SunMarc has tested or deployed Beam.