Current outcome
Reflection AI unveiled Beam on October 5, 2026, its first frontier open-weight sparse MoE model with 501B total and 23B active parameters for coding, reasoning and agentic tasks. New coverage shows Reflection's own benchmarks place Beam at 44.4 on DeepSWE, close to GLM-5.2 but behind the newest Chinese open-weight models. Beam is still in final safety testing with limited early access; full weights, technical report and model card are planned for later in October 2026 under Apache 2.0. Training used 6,144 GB300 GPUs for pretraining under 4 weeks on 23.8 trillion tokens, followed by 4 weeks of reinforcement learning on 10,500 GB300 GPUs generating over 100 million rollouts.
Progress timeline
2 material updates- #01
Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost
Reflection AI unveiled Beam on Monday, October 5, 2026, its first frontier open-weight model. According to the company's blog, Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic workloads. The launch puts a US lab in direct competition with lower-cost Chinese open-weight models such as DeepSeek and Moonshot AI's Kimi in coding and related tasks, a segment where Chinese releases have set the cost benchmark. Reflection is Nvidia-backed and is pitching Beam and future models at enterprises and sovereign nations through an "AI factories" offering that would let institutions train the models on their own proprietary data. The available coverage does not include independent benchmark results or licensing terms.
Source evidence: TechCrunch AI
- #02
Reflection终于交卷：501B模型Beam亮相，仍落后最新中国头部开源模型
New coverage adds Reflection's own benchmark results showing Beam scoring 44.4 on DeepSWE, close to GLM-5.2's 44.0 but behind Qwen 3.8-Max, GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Training details: pretraining on 6,144 GB300 GPUs processing 23.8 trillion tokens in under 4 weeks, followed by 4 weeks of reinforcement learning on 10,500 GB300 GPUs generating over 100 million rollouts. Beam is still in final safety testing with limited early access; full weights, technical report and model card are planned for later this month under Apache 2.0.
State after update: Reflection AI unveiled Beam on October 5, 2026, its first frontier open-weight sparse MoE model with 501B total and 23B active parameters for coding, reasoning and agentic tasks. New coverage shows Reflection's own benchmarks place Beam at 44.4 on DeepSWE, close to GLM-5.2 but behind the newest Chinese open-weight models. Beam is still in final safety testing with limited early access; full weights, technical report and model card are planned for later in October 2026 under Apache 2.0. Training used 6,144 GB300 GPUs for pretraining under 4 weeks on 23.8 trillion tokens, followed by 4 weeks of reinforcement learning on 10,500 GB300 GPUs generating over 100 million rollouts.
Source evidence: theblockbeats