Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost

1. #0110.05
   Initial2026-10-05 19:33 · publication time

   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

2. #0210.06
   Confirmation2026-10-06 13:05 · publication time

   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

3. #0310.16
   Confirmation2026-10-16 14:58 · publication time

   英伟达传洽购的Reflection开源Beam权重：美国开源第二，落后Meta 9分

   Reflection has uploaded the weights of Beam 1.0 to Hugging Face, fulfilling its October 5 preview, under Apache 2.0 with BF16, FP8 and NVFP4 versions and 1M-token context. It also open-sourced the Mirror CLI coding agent framework and DFlash2 speculative decoding accelerator, and worked with vLLM to optimize inference, cutting per-step decoding time for a single request from 8.95 ms to 3.30 ms (2.71x speedup). Artificial Analysis gives Beam a score of 30, ranking second among US open-weight models but behind Chinese models. The release has support from over 40 partners including AWS, Google Cloud and Nvidia. Nvidia's reported talks to further invest in or acquire Reflection are ongoing.

   State after update: Beam 1.0's full weights were released on Hugging Face under Apache 2.0 around October 16, 2026, with BF16, FP8 and NVFP4 versions and 1M-token context. Reflection also open-sourced the Mirror CLI coding agent framework and DFlash2 speculative decoding accelerator, and worked with vLLM to optimize inference. Independent Artificial Analysis evaluation gives Beam a score of 30, ranking second among US open-weight models but behind Chinese models such as DeepSeek and GLM. The release has support from over 40 partners. Nvidia's reported talks to further invest in or acquire Reflection are ongoing.

   Source evidence: theblockbeats

All events · Back to the feed