Reflection Open-Sources Beam-1.0, vLLM Follows Up

Episode 1 · Reflection Open-Sources Beam-1.0: 501B MoE Model with Apache 2.0 and 40+ Partners (2026-10-16, 20 posts)

On October 16, Reflection AI released Beam-1.0, its first open-weight model — a MoE architecture with 501B total parameters and 23B active per token, supporting a 1M token context. It is open-sourced under Apache 2.0, targets coding, reasoning, and agent tasks, and launched alongside 40+ ecosystem partners (Nebius, Ollama, Databricks, AWS, Google Cloud, NVIDIA, and others).

Confirmed

• Model specs: 501B MoE, 23B active per token; 52 layers, 256 experts per layer with 8 routed; 1M token context
• Full collection available on Hugging Face, including the original Beam-1.0-501B-A23B plus FP8 and NVFP4 quantized versions and a DFlash2 version
• The release also includes optimized inference support, quantized weights, and a coding harness
• Hugging Face team benchmarks: SWE-bench at 81.6%, with the NVFP4 quantized version showing almost no performance drop
• Licensed under Apache 2.0, focused on token efficiency for coding and agent use cases

Why it matters

Beam is Reflection AI's first open-weight model, fully open under Apache 2.0 with quantization and deployment tooling included, and backed by 40+ cloud and tooling partners — significantly lowering the barrier to enterprise adoption. HF benchmarks show the NVFP4 quantization is nearly lossless, meaning coding and agent capabilities close to the original can be retained even on low-precision hardware, putting direct competitive pressure on the open-source coding/agent model landscape.

• Reflection launches Beam 1.0: open-weight 501B/23B model with 1M context under Apache 2.0 — bhutanisanyam1 · 2026-10-16
• Reflection ships Beam: 501B MoE open model with 23B active, 1M context, Apache 2.0 — ben_burtenshaw · 2026-10-16
• Beam benchmarks: SWE-bench 81.6%, NVFP4 quantization barely hurts, fits 4×B200 — ben_burtenshaw · 2026-10-16
• Reflection open-sources Beam under Apache 2.0 with 40+ partner ecosystem — bhutanisanyam1 · 2026-10-16
• Beam 1.0 collection lands on Hugging Face with FP8, NVFP4 and DFlash2 variants — victormustar · 2026-10-16
• Reflection releases Beam 1.0, its first open-weight model for coding, reasoning and agents — paf1138 · 2026-10-16
• Reflection's First Open-Weight Model Beam Debuts: 501B-Total MoE With 23B Active, Built for Coding and Agents — baseten · 2026-10-16
• Reflection's Beam: a 501B MoE open model with 1M context, Apache 2.0, agentic focus — mervenoyann · 2026-10-16
• Reflection launches Beam 1.0, its first open-weight model built for coding, reasoning and agentic tasks — tomaarsen · 2026-10-16
• Reflection's Beam lands on Hugging Face: 501B MoE, 1M context, Apache 2.0 open weights — tomaarsen · 2026-10-16
• vLLM ships Day-0 support for Reflection-Beam-1.0, a 501B-param MoE with 1M context — vllm_project · 2026-10-16
• Reflection releases open MoE Beam-1.0: 501B total/23B active, 1M context, 3x fewer reasoning tokens — jacek2023 · 2026-10-16
• Reflection AI open-sources Beam model weights under Apache 2.0 with 40+ partners — victormustar · 2026-10-16
• Reflection AI's 501B open model Beam fine-tunes with LoRA on a single 8x B300 node — QGallouedec · 2026-10-16
• Reflection's Beam, a 501B-parameter open MoE model with 1M context, lands on Hugging Face under Apache 2.0 — QGallouedec · 2026-10-16
• Reflection AI open-sources Beam under Apache 2.0 with 40+ ecosystem partners — zhyncs42 · 2026-10-16
• Ollama Lists Reflection's Beam: 501B/23B MoE With 1M Context and Tool Support — ollama · 2026-10-16
• Reflection open-sources 501B MoE Beam; vLLM day-0 support hits 708 tokens/s — Xianbao_QIAN · 2026-10-16
• Reflection AI open-sources Beam: 501B-param MoE model under Apache 2.0 for agentic coding — arena · 2026-10-16
• Reflection AI open-sources Beam: a 501B-A23B MoE trained on 10K GB300s, Apache 2.0 — AI4Code · 2026-10-16

Episode 2 · vLLM Details Beam Architecture and Long-Context Optimizations (2026-10-16, 3 posts)

vLLM detailed Beam's architecture—52 layers with 256 experts activating 8 per token (23B of 501B params)—and kernel-level long-context optimizations, cutting GB200 forward time by up to 20%.

• Inside Beam: 8-of-256 expert routing across 52 layers, sliding window plus NoPE attention — vllm_project · 2026-10-16
• vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
• Inside Beam: 501B params, only 23B active per token via 8-of-256 expert routing — vllm_project · 2026-10-16