Most of the strongest open AI models people can download and run themselves have come from Chinese labs. A Western rival that claims frontier-level coding skill at a fraction of the inference cost would matter—if the numbers hold up.
On 5 October 2026, Reflection AI published an introduction to Beam, its first open-weight model: a sparse Mixture-of-Experts (MoE) system described as 501 billion total parameters with 23 billion active at once, focused on coding, reasoning and agentic workloads. [1]
What is Beam, in plain terms?
An MoE model is like a large team of specialist sub-models (“experts”) where only a subset wakes up for each token. That design can keep total capacity high while limiting how much compute each answer costs. Reflection says Beam was pretrained on 23.8 trillion tokens, and that a high-compute reinforcement-learning run used 10,500 NVIDIA GB300 GPUs over about four weeks and generated over 100 million rollouts. Those figures are the company’s. [1]
The firm also says midtraining extends effective context to one million tokens. It plans an Apache 2.0 weight release later in October, with early access via a waitlist while red-teaming continues. [1]
How does Reflection score it?
Reflection’s own tables compare Beam to models such as GLM 5.2, Qwen 3.8 Max and Inkling on coding, agentic and reasoning suites. The company claims frontier-level capability at three to four times less inference compute than some larger peers on selected tasks. None of those benchmark or efficiency claims have been independently verified in the coverage cited here. [1]
TechCrunch notes that Reflection has raised about $4.7 billion and positions Beam as a Western open-weight rival to Chinese open models. [2]
What this does not prove
- All capability and efficiency figures are company-reported; third-party evals are still pending.
- Beam is text-only in this preview, so comparisons with multimodal rivals are not like-for-like.
- Weights, a full model card and public safety evaluations are not yet available.
The Bottom Line
Reflection is previewing Beam as a large open-weight MoE aimed at coding and agent work, with weights promised under Apache 2.0 later this month. The efficiency pitch is bold; until outsiders can run the model and re-score the tables, treat the leaderboard claims as the company’s case, not settled fact.
Sources
- Reflection AI, “Introducing Beam,” 5 October 2026 — https://reflection.ai/blog/introducing-beam
- TechCrunch, 5 October 2026 — https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/
