The October 6 post covering Reflection AI’s Beam announcement ended with “whether the weights follow is the story to watch for the rest of October.” They followed. On October 16, Reflection posted weights to Hugging Face under Apache 2.0: a 501B-parameter, 23B-active sparse mixture-of-experts model in BF16, FP8, and NVFP4 checkpoints, alongside a speculative-decoding drafter, reference implementation, and tokenizer. The benchmarks are the same as announced — SWE-Bench Pro v2-Hard at 77.2, Terminal Bench v2.1 at 80.1 — and still vendor-reported.
The more interesting release may be Mirror CLI, the terminal agent harness Reflection used to train and evaluate Beam.
Mirror’s architecture is a Python frontend handling prompts, skills, and tooling over a Rust core that drives the model over a local connection. The Rust layer handles evaluation as well, which Reflection says reduces the delta between their benchmark environment and what developers actually run — a meaningful claim, because many benchmark regressions on released models trace to exactly that gap. The agent harness runs parallel tasks in separate Git worktrees, supports MCP tool connectivity, and lets you view and approve tool calls or steer a running task. The design closely parallels Claude Code and OpenCode 2.0, neither of which open-sources its harness core.
The practical significance of Mirror being open is that you can see the scaffolding Beam was optimized against. When a model underperforms expectations in practice, one frequent cause is misalignment between the agent loop it trained on and the one you’re running it through. Having both the model and the harness from the same source removes that variable, at least as a starting point.
The training scale details from the announcement still deserve attention. Reflection ran more than 100 million RL rollouts across roughly 1.3 billion sandbox environments, averaging 110,000 concurrent rollouts with peaks near 170,000. Policy staleness — the problem that long-horizon rollouts may have been generated by a checkpoint that’s already several updates out of date — is a real technical challenge at this scale, and Reflection says they developed specific algorithms to handle it. The 4-week RL run on 10,500 GB300 GPUs is simply a lot of compute directed at a verifiable reward signal.
The other detail worth noting is the self-distillation path for specialization. Rather than running a full RL sweep to adapt Beam to a specific domain, Reflection says you can generate traces with Beam, keep the verified successes, and train a LoRA adapter on them. This makes domain adaptation accessible to teams without frontier-scale GPU budgets, though how much the approach degrades relative to full RL remains to be measured.
Whether Beam becomes widely used depends more on inference cost than the Apache license. At 501B total parameters with FP8 or NVFP4 quantization, the floor is probably a multi-GPU server; most teams won’t run that locally. But the weights being Apache 2.0 with no use-restriction rider means the barrier is hardware, not legal review. That’s a different situation from “available via API” and matters for anyone building on top of it.