Reflection AI, the Nvidia-backed startup founded by former Google DeepMind researchers, has unveiled Beam, its first open-weight model. Beam is built for coding, reasoning and agentic work. Reflection will release its weights later this month, the company said in a blog post on Monday.
Anyone can download an open-weight model and run it on their own hardware. Beam is still going through final red-teaming and evaluations. For now, a select group of users can try an early version through a waitlist. Axios first reported on Sunday that the launch was close.
What Reflection says Beam can do
Beam is a mixture-of-experts model with 501 billion parameters in total. Only 23 billion of them are active at a time. Reflection pretrained it on 23.8 trillion tokens, and its context window reaches one million tokens.
On coding and agentic tasks, Reflection said Beam is competitive with GLM 5.2 from China’s Z.ai. It is also approaching Alibaba’s Qwen 3.8-Max. Moonshot AI’s Kimi K3 remains ahead on raw capability, it said. Beam’s advantage is efficiency, according to the company.
On advanced reasoning benchmarks, it matches GLM-5.2 while using three to four times less inference compute. Reflection described that figure as an approximate comparison, not a measured cost.
The company’s own table gives Beam 80.1 on Terminal Bench v2.1, against 88.3 for Kimi K3. On SWE-bench Verified, Beam scored 80.9, against 77.6 for Inkling, the open model from Thinking Machines Lab. All the scores come from Reflection, and TNW has not independently verified them.
How it was trained
Reflection said it pretrained Beam in under four weeks on 6,144 Nvidia GB300 GPUs. Its reinforcement learning run then ran for four weeks on 10,500 GB300 GPUs. It produced more than 100 million attempts at tasks. The company believes this is one of the largest such runs by any open lab.
During training, Beam got better at browsing the web even though no browsing tasks were in the mix, Reflection said. Given web access, it learned on its own to query other AI models and to use text-recognition tools to read documents.
Reflection signed compute deals this summer, including a $6.3bn deal with SpaceX and a $1bn deal with Nebius.
Safety and what comes next
Reflection trained a second model from the same base for safety and alignment, then merged the two. Its safety results will appear in Beam’s technical report, it said. It also plans to open-source the safety tests it built internally.
Chief executive Misha Laskin told Semafor that two government bodies are assessing the model with Reflection. They are the US Center for Advancing Innovation and Standards for Super Intelligence and the UK’s AI Safety Institute.
Laskin and co-founder Ioannis Antonoglou discussed the launch on the Sources podcast. Laskin compared closed models to renting an apartment.
“The only way to own intelligence is, by definition, if it’s open,” Laskin said.
Antonoglou was asked whether a model could become too capable to release openly.
“It is possible that you get to a level of capability that you want to just be more careful with how you deploy it,” Antonoglou said.
Reflection will release Beam under an Apache 2.0 licence this month, with a technical report, a model card and tools for running and fine-tuning it. It joins other US open-weight efforts, including Nvidia’s Nemotron models. Reflection said it is already training its next model, and Antonoglou said significantly bigger models will come in 2027.