Reflection AI introduced Beam on Monday, October 5, 2026, its first open-weight model. It is a text-only Mixture-of-Experts with 501 billion total parameters and about 23 billion active per token, built for coding, reasoning, and agentic work.
The weights are not public yet. The company says Beam is in final red-teaming and that it will release weights, a technical report, a model card, and developer artifacts later this month, with an Apache 2.0 license planned. Early access is through a waitlist on Reflection’s platform.
What the company claims
In its official post, Reflection says Beam was pretrained on 23.8 trillion tokens and has an effective 1 million token context window. The reinforcement-learning phase reportedly generated more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks, with about 1.3 billion training and evaluation sandboxes.
The benchmark numbers come from Reflection and do not yet have a full independent check. The company cites 80.9 on SWE-bench Verified, 80.1 on Terminal-Bench v2.1, 65.5 on SWE-bench Pro v1, and 44.4 on DeepSWE v1.1. It also says that, on advanced reasoning, Beam matches Z.ai’s GLM-5.2 while using 3 to 4 times less inference compute. Where larger models such as Kimi K3 remain ahead on raw capability, Reflection’s argument is efficiency.
Artificial Analysis said on X that Reflection granted it access and that it is benchmarking the model independently. The early signal, according to the lab, is high token efficiency for the observed level of intelligence. That is still an early indicator, not a closed ranking.
Why it matters
Beam joins the race for a Western open model that can compete with DeepSeek, Qwen, and GLM without the same inference cost. TechCrunch notes that the startup, founded in 2024 by former Google DeepMind researchers, is positioning itself against closed labs, Chinese open models, and Inkling from Thinking Machines. Beam is text-only; Inkling is multimodal.
Reflection also describes a reasoning-effort setting so users can trade shorter answers for longer chains when the task justifies the cost. Until the weights ship, what changes in practice is the waitlist and the promise of distribution through hyperscalers and neoclouds at launch.
Image credit: Reflection AI
Sources
- Reflection official post: Introducing Beam
- Reflection announcement on X
- Artificial Analysis on independent benchmarking
- TechCrunch
Transparency: This content was created, edited, or reviewed with the help of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Check the original sources for the full context.
By GeekikiBot
