Technology
Reflection AI introduces 501B open-weight model Beam
Startup claims its new model rivals GLM-5.2 at lower compute costs before public release
The short version
- Reflection AI unveiled Beam, an open-weight mixture-of-experts model featuring 501 billion total parameters and 23 billion active parameters.[Hacker News · TechCrunch]
- The startup claims the model matches Z.ai's GLM-5.2 on advanced reasoning while requiring three to four times less inference compute, though this remains unverified.[Hacker News · TechCrunch]
- The model was trained on 23.8 trillion tokens using 10,500 Nvidia GB300 chips over four weeks of reinforcement learning.[Hacker News · TechCrunch]
- Reflection plans to publish Beam's model weights, technical documentation, and developer artifacts later in October 2026.[Hacker News · TechCrunch]
Key facts
- Reflection AI announced Beam, an open-weight sparse mixture-of-experts model containing 501 billion total parameters and 23 billion active parameters.[Hacker News · TechCrunch]
- The model was pretrained on 23.8 trillion tokens and includes a 1 million token context window.[Hacker News · TechCrunch]
- Training involved reinforcement learning on 10,500 Nvidia GB300 GPUs across four weeks, generating over 100 million rollouts.[Hacker News]
- Reflection previously signed agreements valued at over $7 billion with SpaceX and Nebius to access Nvidia GB300 chips through 2029.[TechCrunch]
- Reflection plans to release the model weights, technical documentation, model card, and developer artifacts later in October 2026.[Hacker News · TechCrunch]
What remains uncertain
- Reflection's claims that Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks and requires three to four times less inference compute have not been independently verified.[TechCrunch]
Sources
Outlet counts describe coverage, not independent confirmation. Reports may share a wire service or original source.