Reflection AI dropped Beam—501B total params with 23B active (sparse MoE architecture). Apache 2.0 licensed, so fully open-weight.
Hit 80% on SWE-bench Verified, which puts it in serious territory for code generation and software engineering tasks. For context, that benchmark tests real-world GitHub issue resolution.
The key unlock here: this scale requires distributed training infra. You can't run 501B params on a single node—needs multi-GPU clusters with proper tensor/pipeline parallelism. The fact it's open-weight means anyone with distributed compute can fine-tune or deploy, but the barrier is infrastructure, not licensing.
This is the Western response to closed Chinese frontier models. Sparse activation (23B active) keeps inference costs reasonable while the full 501B parameter count gives it massive capacity during training.
Hit 80% on SWE-bench Verified, which puts it in serious territory for code generation and software engineering tasks. For context, that benchmark tests real-world GitHub issue resolution.
The key unlock here: this scale requires distributed training infra. You can't run 501B params on a single node—needs multi-GPU clusters with proper tensor/pipeline parallelism. The fact it's open-weight means anyone with distributed compute can fine-tune or deploy, but the barrier is infrastructure, not licensing.
This is the Western response to closed Chinese frontier models. Sparse activation (23B active) keeps inference costs reasonable while the full 501B parameter count gives it massive capacity during training.