Reflection launches Beam: 501B open-weight MoE trained on 23.8T tokens with 100M+ RL rollouts
gethackteam · x · 2026-10-06
Reflection announced Beam, its first open-weight model: a sparse MoE with 501B total / 23B active parameters, built for coding, reasoning, and agentic workloads.
Training: pretrained on 23.8T curated tokens (web + licensed proprietary data), matching or beating similar-sized open base models; a high-compute RL run generated 100M+ rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks.
Positioning: advances the Western open-weight frontier on coding and agentic tasks — competitive with GLM 5.2 and approaching Qwen 3.8-Max, though frontier open models like Kimi K3 stay ahead on raw capability; Beam's edge is inference-time efficiency. Benchmark charts cover agentic coding/terminal, reasoning, and tool calling.
Rollout: final red-teaming underway, early access open now; weights, tech report, model card, and developer artifacts ship later this month.
More from Infra
- Nvidia down ~40% from October peak as Anthropic IPO pops 34% above expectations — The Verge AI · 2026-11-20
- Ion Stoica explains why complex AI workloads increasingly fit Ray's design — PyTorch · 2026-11-20
- GPU rents up, token prices down: inference providers' break-even load jumps 22%–75% in 5 months — AccBalanced · 2026-11-20
- Missing config file was throttling GPUDirect Storage — fix yields 3x reads, 9x writes — AccBalanced · 2026-11-20
- NVIDIA Dynamo cuts RL rollout bottlenecks with KV-aware routing and delta weight updates — NVIDIAAI · 2026-11-20
- Arm reaches the AI laptop: Surface Laptop Ultra runs 120B models locally — kimmonismus · 2026-11-20