Reflection announces Beam, a 501B open-weight model it says needs 3 to 4 times less compute
Reflection AI announced Beam on 5 October, its first frontier open-weight model. Beam is a sparse mixture-of-experts model with 501 billion total and 23 billion active parameters, trained on 23.8 trillion tokens, with its context window extended to 1 million tokens. It will be released under the Apache 2.0 licence, with weights, a technical report and a model card due later this month.
Reflection's own results put Beam close to, but mostly behind, the GLM models it compares against: 80.1 on Terminal Bench v2.1 against 81.0 for GLM 5.2, and 77.2 on SWE Bench Pro v2-Hard against 84.3 for GLM 5.3. The company says Beam uses 3 to 4 times less inference compute than GLM-5.2 on reasoning benchmarks. These figures are Reflection's own and have not yet been checked independently.
If the efficiency claim holds, an open model near the top of the open field at a third of the compute would cut the cost of serving reasoning-heavy agents. Wait for the weights and run your own tests before planning around it.
Drafted with AI from the original source, then checked and edited by Sridhar Mukkandi. Spotted a mistake? Write to hello@tarkika.com and we'll correct it here.