What changed
On October 16, 2026 Reflection released Beam 1.0 model weights on Hugging Face under Apache 2.0, with BF16, FP8 and NVFP4 variants. Its model card specifies 501.3 billion total parameters, 23.4 billion active per token, a one-million-token context window and text-only input/output. Reflection also launched hosted API access and reference agent/inference integrations. The vendor publishes benchmark tables and a safety technical report, but independent evaluations of released checkpoints remain limited at launch.
Why it matters
The licence allows commercial self-hosting, fine-tuning and integration without the separate model-as-a-service restrictions found in some other recent large-model releases. Open checkpoints allow teams to inspect and benchmark their own agent workloads and avoid an exclusive hosted provider, although hundreds of billions of stored parameters and long-context serving remain beyond ordinary developer hardware. Model usefulness depends on real task quality, latency, memory and cost, not advertised benchmark rank alone.
Open weights, not a small desktop model
Beam is a sparse mixture-of-experts text model with 501.3B total and 23.4B active parameters, 52 layers, 256 experts and a 1M context window. Hugging Face hosts BF16, FP8 and NVFP4 weights. Sparsity reduces active computation per token but does not eliminate the memory needed to store or distribute a very large checkpoint. Self-hosting requires serious GPU resources and validated serving configuration.
Apache 2.0 changes the commercial options
Reflection explicitly releases the weights under Apache 2.0 and says developers may customise, quantise, fine-tune and use outputs for model training. That contrasts with restricted licences on some other large open-weight models. Apache 2.0 obligations still apply, and teams must assess privacy, model-use policy and jurisdictional constraints for their own deployment.
Benchmark claims need independent repetition
Reflection's model card reports 81.6% on SWE-bench Verified for BF16, 80.1% on Terminal Bench v2.1 and 76.5% on SWE Bench Pro v2-Hard, with temperature 0.7 and top_p 0.99. Those are vendor-reported results, not independently established cross-provider rankings. Harness configuration, data contamination controls, inference cost and actual tool reliability may change the comparison.
Agent ecosystem and safety boundaries
Reflection advertises more than 40 partner integrations, hosted access and published safety methodology. The model is text-only, not multimodal. A model card and safety report improve inspectability but do not prove agent sandbox safety or immunity to prompt injection; operators should run their own evaluations and permission controls.