AI Innovation Digest

Open coding-model releases and midtraining recipes raise the benchmark bar

Open coding-model releases and midtraining recipes raise the benchmark bar

Reflection's Apache-licensed Beam 1.0 is a 501B MoE with 23.4B active parameters and a company-reported 81.6% SWE-bench Verified score. Its NVFP4 model card reports broadly similar results to BF16 with some benchmark dips; Nvidia reports coding gains from large-scale software-engineering midtraining. Independent replication, deployment cost, and transfer remain unresolved.

Sources (2)
Updated Oct 21, 2026