Reflection Ships Beam: a 501-Billion-Parameter Apache 2.0 Agent Model With 40-Plus Day-Zero Partners

Reflection AI has made good on its promise. The San Francisco lab, which announced Beam on October 5 with a pledge to open the weights later in the month, published them on October 16 at around 13:00 UTC under a plain Apache 2.0 license — no gate, no attribution riders. The Hugging Face repository, Beam-1.0-501B-A23B, picked up hundreds of downloads within hours, and community quantizers had a GGUF build out in under an hour.

Beam is a 501-billion-parameter mixture-of-experts model with roughly 23.4 billion active parameters per token, a 1-million-token context window, and text-only input. Weights are available in BF16, FP8 and NVFP4, with a speculative-decoding drafter alongside. The architecture uses gated GQA plus sliding-window attention and a modified Harmony chat template, with a June 30 knowledge cutoff.

The pitch is efficiency rather than raw scale. Reflection says Beam reaches solution quality comparable to leading open models using three to four times fewer reasoning tokens — the company's claim, not an independent benchmark. Its own model card shows DeepSWE v1.1 at 43.3 percent in BF16, Terminal-Bench 2.1 at 80.1 and SWE-Bench Verified at 81.6, and it concedes Kimi K3 still leads on DeepSWE. A table published a day earlier by Meta put Beam at 78.7 on MCP-Atlas against Muse Spark Open's 89.2.

The distribution is the real product. More than 40 platforms offered day-zero support, including AWS, Azure, Google Cloud, Oracle, Databricks, Cloudflare, CoreWeave, Baseten, Fireworks, Together, Modal, Nebius, NVIDIA NIM, Ollama, LM Studio, OpenCode, OpenRouter, Vercel and Weights and Biases, plus fine-tuning paths through NeMo RL, TRL, SkyRL, Axolotl and Miles. Google Cloud published day-zero recipes pairing Beam with llm-d and NVIDIA Dynamo, reporting up to 5 times higher throughput and 8 times lower latency on GB300 NVL72 instances — vendor benchmarking, not an independent audit.

Reflection also open-sourced Mirror CLI, the harness Beam was trained with: a Python front end for prompts, skills and tooling driving a Rust agent core, with worktree-isolated parallel agents, resumable sessions, MCP connections and Python extensions. The team says Mirror began as an internal tool so agents could keep working after engineers closed their laptops, with the biggest gains showing up in code review.

There is a geopolitical subtext the company does not hide. Reflection lobbied Washington in May about Chinese open models as a national security concern, and it now sells sovereign deployment to the buyers who care about exactly that — an American open-weights alternative with permissive licensing. With Meta's gated Muse Spark Open, StepFun's conditional license and Reflection's plain Apache all landing within two days, the license column has become part of the open-model scoreboard.