On October 16, Reflection AI put the Beam 1.0 weights on Hugging Face: a 501B-total, 23B-active sparse MoE under Apache 2.0, released in BF16, FP8, and NVFP4 numerics alongside a speculative decoding drafter. Eleven days after the announcement, the startup's first open-weight model has moved from preview to downloadable and self-hostable.

Architecture: a 1M-context sparse MoE

According to the model card, Beam has 501.3B total and 23.4B active parameters; of its 52 layers, 3 are dense and 49 are MoE, with 256 experts. Attention combines gated GQA with sliding windows, the context window is 1M tokens, vocabulary is 152k, the chat template is a modified Harmony format, and the knowledge cutoff is June 30, 2026. It is text-only, positioned for coding, reasoning, and long-horizon agentic workloads.

Token efficiency is listed as the first capability: the card claims Beam reaches solution quality comparable to leading open models using approximately 3x fewer inference-time reasoning tokens. For token-billed agent workloads, that matters more to cost than leaderboard rank.

Benchmarks: narrow gaps across numerics

All evaluations ran at temp=0.7, top_p=0.99. Representative BF16 numbers: SWE Bench Verified 81.6%, SWEBench Multilingual 78.3%, Terminal Bench v2.1 80.1%, AIME 2026 97.7%, GPQA Diamond 90.6%, and LiveCodeBench 88.3%. Notably, lower precision barely hurts: on SWE Bench Pro v2-Hard, the NVFP4 build scores 78.6%, above BF16's 76.5%; across the table, gaps between the three numerics mostly stay within about two points.

Weak spots are printed in the same table: CriPT AA at 16.3%, HLE without tools at 36.0%, Tau3 Banking at 36.1%, and an AA Omniscience index of 10.8. On safety, the harmful-request refusal/pivot rate is 88.9%.

Deployment: the model is the kernel

The real theme of the release is what Reflection calls an open operating system for agents: more than 40 partners cover inference, quantization, fine-tuning, RL, agent harnesses, evaluation, and observability. Day-zero support spans Transformers, vLLM, SGLang, Unsloth, Axolotl, and NVIDIA Dynamo; AWS, Google Cloud, Cloudflare, Fireworks, Together, OpenRouter, Ollama, LM Studio, and Microsoft Foundry offer hosting, while Dell and HPE handle on-premises deployments.

The companion Mirror CLI is a terminal agent harness with a Python frontend over a Rust core, supporting resumable sessions and parallel worktrees. Training-scale numbers made it into the post as well: Beam's RL training ran roughly 1.3 billion sandboxes, peaking at 883,006 vCPUs across 23 clusters on GCP and AWS. Self-distillation recipes are promised in the coming days.

What to make of it

On the plus side, Apache 2.0 plus an explicit no-ban stance on distillation and output training is about as permissive as open weights get today, and shipping three numerics at once spares the community its first wave of quantization stumbles. The caveat: an open agent OS is still mostly a coalition narrative — the depth of those 40 integrations will take time to verify, and the 10-30-point benchmarks in the card are a reminder that long-horizon reliability remains the hard part.

Teams that want to self-host can now go from weights to a production endpoint without writing plumbing; observers should watch whether community fine-tunes and third-party evaluations pick up the ball in the next 30 days.

Source: https://reflection.ai/blog/beam-is-available
Model card: https://huggingface.co/reflection/Beam-1.0-501B-A23B