Reflection ships Beam's weights: a 501B Apache-2.0 coding model
The first US open model at this tier runs free on Ollama today, fits a single 8-GPU box, and claims GLM-5.2 quality at 3–4x less inference compute.
Copy markdown
501B open weights, finally downloadable
Beam is a sparse MoE — 501B total, 23B active, 1M-token context — released under Apache 2.0, so you can self-host, fine-tune, and ship it commercially with no strings. It posts 80.9% on SWE-bench Verified and 80.1% on Terminal-Bench v2.1, the first US open model to land in that range.
Run it free this weekend, zero setup
Ollama is a day-zero partner: `ollama run beam:cloud` gets you the model free for a limited time with zero data retention, and `ollama launch claude --model beam:cloud` wires it straight into Claude Code, Codex, or OpenCode. Reflection's own hosted beta (model `Beam-501B-A23B`, OpenAI-compatible at api.reflection.ai) is free during the beta too.
Fits on a single 8-GPU box
The 8-bit and 4-bit builds run on one eight-GPU server (H100/H200/GB300); full precision wants tens of GPUs. Reflection ships one-command Kubernetes containers plus Mirror CLI, an open-source coding agent, with vLLM MXFP4 serving on the way — so on-prem and air-gapped deploys are realistic, not aspirational.
The pitch: fewer tokens, not cheaper tokens
Reflection claims Beam matches GLM-5.2 on reasoning while burning 3–4x less inference compute — roughly 25K tokens per Terminal-Bench attempt versus ~80K for GLM. If that holds on your prompts, the real saving is lower token spend per finished task, not a lower per-token rate (none is published yet).
The catch: vendor benchmarks, and China still leads
Every number is Reflection's own and unverified. On Terminal-Bench v2.1 its 80.1 trails DeepSeek V4.1 Flash (90.6) and Kimi K3 (88.3), and its DeepSWE agentic-coding score (44.4) sits behind both. The hosted beta also caps context at 262K tokens, far under the 1M the weights support. Treat Beam as a cheap-to-run workhorse, not a frontier replacement.