Reflection's Beam is now out as open weights under Apache 2.0. It is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, 52 layers, interleaved local and global attention, pretrained on 23.8 trillion tokens. Midtraining extends the effective context to 1 million tokens; reinforcement learning ran at a 256K maximum.
Read the benchmark table honestly and Beam is not the strongest open model available. On Terminal-Bench v2.1 it scores 80.1, against GLM-5.2 at 81.0, Qwen 3.8 Max at 86.6, GLM-5.3 at 88.2, Kimi K3 at 88.3 and DeepSeek V4.1 Flash at 90.6. It also posts SWE-Bench Verified 80.9, SWE-Bench Pro v1 65.5 and DeepSWE v1.1 44.4.
The claim is efficiency, not placement. Reflection reports reasoning scores comparable to GLM-5.2 while using three to four times less inference compute — roughly 25,000 tokens per Terminal-Bench attempt against GLM-5.2's 80,000. Those figures are the company's own estimates, calculated as 2 x active parameters x mean generated tokens, and they exclude prompt prefill, attention and serving overhead. Treat them as a direction, not a measurement.
Who this is for
If you run agent loops where the same task is attempted repeatedly, token count is your bill and a model that reaches a similar answer in a third of the tokens is worth more than four points of benchmark. If you need the highest score available, DeepSeek V4.1 Flash is ten points ahead. Training used 6,144 Nvidia GB300 NVL72 GPUs for pretraining in under four weeks, and 10,500 GB300s over four more for RL across 100 million rollouts.
Sources:
Reflection ·
MarkTechPost