Reflection AI announces Beam, a 501B open-weight model, with Apache 2.0 weights promised this month

Reflection AI announced Beam today, its first open-weight model: a sparse Mixture-of-Experts with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic work. It is text-only. The weights are not out yet. Reflection says it will release them under Apache 2.0 later this month, with a technical report and model card. Until then, access is an early-access waitlist.

WHAT REFLECTION SAYS IT SCORES

These are Reflection's own figures, not independent results.

- SWE-bench Verified: 80.9% - Terminal-Bench 2.1: 80.1% - GPQA Diamond: 90.5% - Humanity's Last Exam, no tools: 36.2% - AIME 2026: 97.8%

Reflection's own table does not put it on top. On Terminal-Bench 2.1, Kimi K3 (88.3%) and DeepSeek V4.1 Flash (90.6%) lead it and GLM 5.2 sits level (81.0%). Reflection's pitch is efficiency instead: it claims reasoning scores comparable to GLM-5.2 at 3 to 4 times less inference compute. That is an estimate from active parameters times generated tokens, not a measured serving cost.

HOW IT WAS TRAINED

Per Reflection: pretrained on 23.8 trillion tokens, then a reinforcement learning run of more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks. It says midtraining extends the effective context to 1M tokens.

WHAT WE DO NOT KNOW

There are no weights, no API price and no public endpoint yet, so nobody outside Reflection has run it. "This month" is Reflection's own window. Beam is now on our watchlist, and it moves to live when a Beam id appears on Hugging Face or in a catalog we sweep.