Reflection, a US-based AI lab, has announced its new open-weight model named Beam. This model features a 501 billion total parameter, 23 billion active parameter Mixture-of-Experts architecture, trained from scratch with a focus on coding, agentic, and scientific tasks. Beam was trained on 23.8 trillion tokens and underwent a significant reinforcement learning phase using 10.5K NVIDIA GB300 GPUs over four weeks, resulting in over 100 million rollouts. While competitive with models like GLM 5.2 and approaching Qwen 3.8-Max in some areas, it is generally placed below frontier models like Kimi K3 and DeepSeek V4.1 Flash, though it offers notable inference efficiency. AI
IMPACT Sets a new benchmark for open-weight model efficiency and training scale, potentially influencing future open-source development.
RANK_REASON Frontier-lab model release with system card and technical details.
Read on Mastodon — sigmoid.social →
- Beam
- GLM 5.2
- Kimi k3
- NVIDIA GB300
- Reflection
- Agr-flash
- Aleph Alpha Kolibri
- Artificial Analysis
- Axios
- Command Code
- DeepSeek V3
- DeepSeek V4.1 Flash
- Qwen 3.8 Max
- Reka Rho-1
- Solar Mini 4
- Upstage
AI-generated summary · Google Gemini · from 6 sources. How we write summaries →
