Reflection AI Releases Beam 501B Open Weights Under Apache 2.0
Reflection AI shipped the weights for Beam, its 501-billion-parameter sparse MoE coding and agent model (23B active), under an Apache 2.0 license on Hugging Face, with BF16, FP8 and NVFP4 builds plus a DFlash2 drafter—ending the Oct. 5 “weights later this month” wait.
By 1EyeAI Desk ·
On Oct. 16, 2026 at about 9:01 AM ET, Reflection posted that Beam is now available across platforms with open weights under the Apache 2.0 license, with what it calls an ecosystem of 40+ partners. The company says it is releasing weights, tokenizer and a reference implementation; BF16, FP8 and NVFP4 numerics; and a DFlash2 drafter. Hugging Face now lists reflection/Beam-1.0-501B-A23B (updated within the hour of this note), plus FP8, NVFP4 and DFlash2 variants. Ollama lists beam:cloud for cloud inference.
Beam was first announced Oct. 5 as a sparse mixture-of-experts model with 501B total / 23B active parameters, a 1M-token context window, and a focus on coding, reasoning and agentic workloads. Reflection’s Oct. 5 blog said weights would come “later this month” after red-teaming; today’s posts are the ship. The Hugging Face model card reports 501.3B total / 23.4B active, 52 layers, 256 experts, knowledge cutoff June 30, 2026, and text-only input/output.
Benchmark figures on the model card are company-reported (examples: SWE-bench Verified 81.6% BF16; Terminal Bench v2.1 80.1% BF16). Treat them as vendor numbers until independent leaderboards catch up. Reflection also claims roughly 3× fewer inference-time reasoning tokens for comparable quality versus leading open models—again a company claim. Self-hosting a 501B MoE still needs serious multi-GPU memory even with 23B active; most people will use hosted or partner endpoints first.
What this means for you: Developers can pull weights from Hugging Face or try ollama run beam:cloud. Run your own coding and agent tasks before trusting the table scores. Everyday chatbot users: nothing required unless you want an Apache-licensed open coding model in your stack.