Reflection debuts Beam: a 501B open-weight model for coding
The first US startup open model to rival top Chinese weights—23B active, Apache 2.0, early access now, weights this month. Plus Cloudflare's Web Search API.
Copy markdown
A frontier-class open model, Apache 2.0
Reflection's Beam is a 501B-parameter sparse MoE with just 23B active, pretrained on 23.8T tokens for coding, reasoning, and agent workloads. It's the first US-startup open model to match or beat the top Chinese open weights—and it's Apache 2.0, so you can self-host, fine-tune, and ship commercially with no strings.
Beats GLM-5.2 on a quarter of the compute
Beam scores 80.9 on SWE-bench Verified and 80.1 on Terminal-Bench v2.1, matching GLM-5.2's reasoning while using 3–4× less inference compute—despite GLM-5.2 carrying ~250B more parameters. With only 23B active and a 256K context (1M in midtraining), near-frontier coding finally fits on hardware you can actually rent.
Early access now; weights land this month
Sign up at platform.reflection.ai today while Beam finishes red-teaming; the weights, technical report, model card, and fine-tuning artifacts drop 'later this month' via open-source partners. If you've wanted a US open-weight base you can legally build a product on, this is the one to grab the moment it ships.
Cloudflare gives your agents live web search
Cloudflare's Web Search API (open beta) lets agents query the live web and get back clean title/URL/snippet results through AI Gateway, across Exa, Ceramic, and Linkup—billed at each provider's list price with no markup. Swap providers without code changes and skip building your own scraping layer.