Reflection公司正式推出其首个开源大模型Beam，这是一个采用稀疏混合专家架构的模型，旨在为编码、推理和智能体工作负载提供高效解决方案1。根据来源披露，Beam拥有501亿总参数，其中230亿为活跃参数1，在23.8万亿高质量Token上进行了预训练，并配备100万Token的上下文窗口2。
在性能表现上，Beam在推理基准测试中与GLM-5.2性能相当2，同时在推理计算成本上实现了显著优势——相比GLM-5.2节省3至4倍计算资源2。该模型的训练规模庞大，Reflection在10.5K块NVIDIA GB300 GPU上进行了为期4周的强化学习训练，产生了超过1亿次rollouts1。
Reflection已从英伟达、Sequoia Capital和Lightspeed Venture Partners等投资者筹集约47亿美元2，最后一轮融资的估值达250亿美元2。公司今年夏天与SpaceX和Nebius签署的集体协议价值超过70亿美元，确保了至2029年对英伟达GB300芯片的接入权2。Beam将采用Apache 2.0许可证开源，公司计划本月发布模型权重、技术报告、模型卡和开发工具12。
Reflection AI has unveiled Beam, its first open-source frontier model designed to compete with leading Chinese AI systems while dramatically reducing inference costs.2 The model is a sparse mixture-of-experts architecture with 501 billion total parameters and 23 billion active parameters, trained on 23.8 trillion high-quality tokens.12 Beam achieves performance comparable to GLM-5.2 and Qwen 3.8-Max on reasoning benchmarks while requiring 3–4 times lower inference compute.12
The model was developed through large-scale pretraining combined with reinforcement learning conducted over four weeks on 10,500 NVIDIA GB300 GPUs, generating over 100 million rollouts.1 Beam features a context window of one million tokens and is purpose-built for coding, reasoning, and agentic tasks including tool use and web search.12 Reflection plans to release the model weights, technical report, model card, and developer tools this month under the Apache 2.0 license.1
Reflection has secured approximately $4.7 billion from investors including NVIDIA and Sequoia Capital, with its latest funding round valuing the company at $25 billion pre-money.2 In summer, the company signed collective agreements worth over $7 billion with SpaceX and Nebius granting access to NVIDIA GB300 chips through 2029.2 The company aims to build customized "AI factory" systems for enterprises and sovereign nations.2
评论
还没有评论，欢迎留下第一条。