Reflection AI releases Beam's open weights under Apache 2.0 — Google Cloud adds day-0 support

#What happened

Reflection AI released the open weights for Beam today (Oct 16, 2026), its 501-billion-parameter mixture-of-experts model built for software engineering and long-horizon agentic tasks, under the permissive Apache 2.0 license. Checkpoints ship in BF16, FP8, and NVFP4 at launch. Beam runs 23B parameters per token and carries a 1M-token context. Reflection also open-sourced Mirror, the terminal harness it trained Beam alongside, which runs parallel agents in separate Git worktrees.

Google Cloud announced day-0 support: Beam is available on Google Cloud AI Hypercomputer with ready-to-run deployment recipes. Reflection's VP of Product Joe Spisak says the optimized serving stack (llm-d/vLLM, NVIDIA Dynamo/SGLang) delivered up to 5x higher throughput and 8x lower latency on A4X Max (NVIDIA GB300 NVL72) instances in their tests; G4 VMs (NVIDIA RTX PRO 6000) claim up to 2x throughput.

First independent numbers are in from Artificial Analysis: Beam scores 1.5 on the Intelligence Index net score, ahead of K2 Horizon (-3.0) and Qwen3.8 27B (-10.0), but trails Meta's Muse Spark Open (39) and leads NVIDIA's Nemotron 3.5 Super VL (26) among open-weight models released this week. On the Coding Agent Index it scores 28 running in Mirror — 15 points behind the Claude Code + Qwen3.8 Max setup (43) — though it is token-frugal, using ~26k output tokens per Intelligence Index task. It leads US open-weight models tested on the Cyber Index at 34% task completion.

#Why it matters

Beam's weight release converts the biggest open-weight efficiency pitch of the month into something developers can actually download. The Apache 2.0 license is the least restrictive of this week's open-model wave (Meta's Muse Spark Open limits commercial use under the Muse Model License), and day-0 cloud recipes plus an open-source agent harness lower the barrier to running long coding-agent sessions. Reflection's own benchmark pitch remains efficiency, not raw parity: it reports 80.9% on SWE-bench Verified for Beam (company-reported, not independently validated) and frames it at 3-4x less inference compute than GLM-5.2. Note Artificial Analysis also flags a 47% hallucination rate in its testing.

Verified Oct 16, 2026. Sources: Google Cloud Blog "Day-0 support with 5x perf for Reflection Beam on Google Cloud" (primary) and Artificial Analysis "Reflection has released Beam, their first model launch" (independent benchmarks). The 80.9% SWE-bench figure is company-reported.