Reflection Beam Joins Agent Arena and Code Arena WebDev as Acquisition Talks Continue
Reflection AI’s Beam model — 501 billion parameters, 23 billion active — joined Arena.ai’s Agent Arena and Code Arena WebDev leaderboards on October 16. The debut comes ten days after the model’s public launch and one week after reports that Nvidia is in acquisition talks with the company.
What Beam Brings to Arena
Beam launched October 6 with vendor-reported SWE-bench Verified at 80.9% and Terminal-Bench 2.1 at 80.1%, matching Claude Opus 4.6 and Gemini 3.1 Pro Preview on the terminal benchmark. DeepSWE v1.1 came in at 44.4%, a gap independent reviewers attributed to harness differences rather than underlying capability.
Arena battle data provides a different signal: preference-weighted quality across real user interactions, blind to benchmark scaffolding. Agent Arena concentrates on multi-step task performance; Code Arena WebDev focuses on web development output quality against human judges.
SWE-bench Verified places Beam at 80.9%, comparable to Gemini 3.1 Pro Preview (80.6%) and Claude Opus 4.6 (80.8%) on the same bash-only harness — and well above the 70–75% range that characterised frontier models six months ago.
The Acquisition Context
Nvidia entered discussions to acquire Reflection AI in early October. At $7.3 billion in committed compute — including a $1 billion Nebius deal announced in July — Reflection has one of the largest infrastructure positions of any post-Series A lab.
Training Beam required 10,500 Nvidia GB300 GPUs over a four-week reinforcement learning run. If the acquisition closes, Beam becomes the only frontier open-weight model directly controlled by a chip manufacturer.
Open Weight Status
Beam carries an Apache 2.0 licence, pending red-team clearance. Weights are not yet publicly released. Once cleared, Beam joins Kimi K3 (2.8T parameters) and Inkling (975B) as one of the largest downloadable models at the frontier tier.
The Arena debut signals that Reflection is treating Beam as a production-quality deployment candidate, not a benchmark-only demonstration. Arena ELO from battle data will provide the first independent performance signal not derived from vendor-administered harnesses.