Reflection open-sources 500B Beam model; Google Cloud day-0 support with 5x throughput
rseroter · x · 2026-10-17
Reflection released Beam, its first open-weight model (Apache 2.0, BF16/FP8/NVFP4 checkpoints), targeting frontier software engineering and long-horizon agentic tasks, claiming 3-4x reasoning compute efficiency vs comparable open models in the 500B class.
Google Cloud announced day-0 support with ready-to-run recipes: serving stack optimized with llm-d (vLLM) and NVIDIA Dynamo (SGLang), delivering up to 5x throughput and 8x lower latency on A4X Max (GB300 NVL72) instances, 2x throughput on G4 VMs (RTX PRO 6000). The architecture uses KV-cache-aware routing, prefill/decode disaggregation, and tiered KV cache across accelerator memory, CPU memory, and storage.
More from Infra
- Litespark serves Nemotron 120B at 962 tokens/sec per user on just two H200 GPUs — NVIDIAAI · 2026-10-17
- SF Signs Data Center Moratorium; Investor Kwok Asks "What Are We Even Doing" — kevinakwok · 2026-10-17
- Inferact, from the creators of vLLM, launches controllable frontier inference platform — Xianbao_QIAN · 2026-10-17
- AMD MI355X beats NVIDIA B300 by 24.4% throughput serving Kimi K3, 87% cheaper per token — vllm_project · 2026-10-17
- Anthropic: Claude Code responds 3x faster, Claude API 1.5x faster via gzip and infra cuts — ClaudeDevs · 2026-10-17
- Reddit user pairs Mac mini M5 Pro with old M4 as sub-agent for local coding agents — Calm-Set6265 · 2026-10-17