16
votes
1K
seen
Reflection AI’s Beam activates just 23B of its 501B parameters
Reflection AI has introduced Beam, an open-weight model aimed at coding, reasoning, and agentic software tasks. It uses a mixture-of-experts design: 501 billion parameters exist in total, but only 23 billion activate for each query, keeping the compute load far below the headline number.
The company says Beam delivers 3-4× better inference efficiency than comparable open models. Its reported scores are 80.1 on Terminal-Bench 2.1 and 80.9 on SWE-bench Verified, with full weights scheduled for release this month.
The company says Beam delivers 3-4× better inference efficiency than comparable open models. Its reported scores are 80.1 on Terminal-Bench 2.1 and 80.9 on SWE-bench Verified, with full weights scheduled for release this month.
yesterday
Reflection AI was reported to be preparing its first open-weight model.
yesterday
Reflection AI officially introduced Beam with 501B total and 23B active parameters.
N
nvoss••17 hours ago
beam's 23b-active MoE is the useful part, but the evidence is still company-provided: 80.9 on SWE-bench Verified and 80.1 on Terminal-Bench 2.1, with no independent reproduction from outside teams or production workloads. if those numbers hold, cheap inference matters.
1K
N
nvoss••17 hours ago
beam's 23b-active MoE is the useful part, but the evidence is still company-provided: 80.9 on SWE-bench Verified and 80.1 on Terminal-Bench 2.1, with no independent reproduction from outside teams or production workloads. if those numbers hold, cheap inference matters.
