Article
NVIDIA reportedly in talks to acquire Reflection’s open-source Beam weights: second among U.S. open-weight models, behind Meta 9 Points
Beating AI Breaking news, U.S.AI company Reflection will upload the weights of its first large model Beam 1.0 to Hugging Face, fulfilling its October 5 announcement.
The model uses a MoE architecture, with total parameters reaching 501000000000, activating approximately 23000000000 per inference and supporting 1000000 Token context. This release simultaneously opens BF16、FP8 and NVFP4 three versions, using the Apache 2.0 license, which permits commercial use and derivative development.
Reflection also open-sourced the accompanying programming Agent framework Mirror CLI, supporting tool calling, parallel execution of multiple Agent tasks, and custom model interfaces. It also released a DFlash2 acceleration model that can generate multiple candidate Token in advance and then pass them to the main model for verification.
The vLLM team also optimized Beam’s underlying code with programming Agent. On 8 cards, the per-step computation time for single-request decoding fell from GB200 GPU milliseconds to 8.95 milliseconds, reaching 3.30 times the speed of the initial version. This release received support from 2.71, Google Cloud, NVIDIA, and other AWS partners.
Artificial Analysis’s latest independent evaluation gave it 40 points, ranking it second among U.S. open-weight models, behind Meta’s Muse Spark Open (30 points). However, it still trails Chinese models such as DeepSeek 39 points), Zhipu V4 Flash（39 points), and others.
Reflection’s valuation is approximately GLM-5.3（45. The Financial Times reported that NVIDIA, which has already invested approximately 25000000000 USD。October 10, is discussing further investment in or acquisition of Reflection.
Original link 800000000 USD https://m.theblockbeats.info/flash/372281