Beam is a frontier-scale coding model under a real open-source license

Reflection, a new US lab, has released Beam, an open-weight language model built for coding and agent work, and the license is the headline. Beam ships under Apache 2.0, a genuine open-source grant, which sets it apart from the restricted terms that labs like Meta attach to their open-weight releases. It is a large model: 501 billion total parameters with 23 billion active per token, a mixture-of-experts design with 256 experts per layer and eight routed at a time, and a 1 million token context window.

The reported benchmarks put it in frontier company for its intended use. Reflection claims 81.6% on SWE-bench Verified and 80.1% on Terminal Bench v2.1, both coding and agent tests, plus 97.7% on AIME 2026 and 90.6% on GPQA Diamond. The team also made choices to keep it cheap to serve. Most layers use sliding-window attention over the previous 512 tokens, with full attention only every sixth layer and at the end, which cuts the KV cache to about 36 KiB per token instead of 208 KiB. The quantized NVFP4 checkpoint is around 280 GB and runs natively on four NVIDIA B200 GPUs, and Reflection says the quantized versions keep most of the full-precision quality.

There is day-zero support in vLLM, SGLang, and Transformers, plus a Rust harness. One demo fine-tuned a small incident-triage model that reached 76.25% accuracy against 4.69% zero-shot, with an 11.9 times latency speedup, a reminder that the weights are yours to specialize.

Why it matters

If you want a strong coding and agent model you can run, inspect, and modify on your own hardware with no license strings attached, Beam is the first at this scale under Apache 2.0. Check whether four B200s or the memory for the larger checkpoints fits your budget before you plan around it.