Reflection 发布开放模型 Beam

Chetan Tekur @chetantekur原推发布于 2026.10.06

在 X 上打开原推

We announced Beam today. I joined Reflection to help teams own the intelligence they put into production. We're building a capable Western open model and the stack to run and customize it. Open models are gaining usage on OpenRouter and Vercel. On Vercel's AI Gateway, open weights grew from 7% of token volume in December to 56% in August, while accounting for 14% of estimated spend. Teams are already moving real work onto these models. https://openrouter.ai/blog/insights/deepseek-v4-adoption/ https://vercel.com/blog/ai-gateway-production-index-september-2026 For agents, efficiency determines how much work you can afford to hand over. A coding agent may call the model many times as it reads a repo, makes changes, runs tests and works through failures. Beam activates 23B of its 501B parameters per token, and we trained it to reward successful solutions while discouraging unnecessary tokens. Lower compute per token and fewer tokens to complete a task can make sustained agent work more affordable. Artificial Analysis is independently benchmarking Beam. Their early indicators suggest it could be among the most token-efficient open models they've seen for its level of intelligence. That's a promising combination for production. https://x.com/ArtificialAnlys/status/2107219177132155233 Owning the intelligence means your team's expertise can become part of how the model works. You can customize it around your repos and workflows, evaluate improvements on your own tasks and keep control of the result. That was a big part of why I joined Reflection. https://chetantekur.substack.com/p/why-i-joined-reflection-ai Sovereignty matters here too. You decide where your data goes, which model version you run and when you upgrade. Beam gives enterprises and governments another Western open option when deployment control and model provenance matter. Openness also gives users more control over safety. Teams can test locally, reproduce failures and adapt safeguards to their deployment. We'll open-source our Beam safety evaluations so the community can test its behavior and improve the evaluations with us. The software around the model matters just as much in daily use. The harness controls context, tools and permissions, which affect both quality and cost. I wrote about this in The Model Interface, and it shapes how we're building the open stack. https://chetantekur.substack.com/p/the-model-interface-issue-11 That stack will let users run, evaluate and customize Beam. It will also include a coding harness designed to put Beam to work in your repos, with your tools. I'm looking forward to sharing more about what we've been building there. Beam is in preview, with Apache 2.0 weights and the open stack planned for later this month. You can sign up for early access through the launch post. What are you looking to try with Beam? https://reflection.ai/blog/introducing-beam

这条是引用着另一条写的：x.com/artificialanlys/status/2107219177132155233