The open-weight club has a new member. Reflection has published the technical report for Beam, which it describes as its first open-weight model: a sparse Mixture-of-Experts with 501.3B total and 23.4B active parameters, 52 layers, 256 experts, and a 152K vocabulary, pretrained on 23.8T tokens and built for coding, reasoning, and agentic workloads (https://arxiv.org/abs/2611.26358). The most informative part of the report is not the benchmark table. It is the compute ledger of the reinforcement learning campaign.
Pretraining: 21.9 days on 6,144 GB300s
Pretraining ran 21.9 days on 6,144 NVIDIA GB300 GPUs, with 21% MFU during steady-state training. The more telling number is goodput, the training progress actually retained after failures and restarts: 77.0% end to end, climbing to 92.3% for the final 20% of steps after stability fixes. For anyone pricing large clusters, goodput is a more honest utilization metric than MFU. The recipe is old-school predictable scaling: model configurations from roughly 300M to 6B validated on a ladder before the 23.4B-active target tier, with the paper claiming validation loss scaled predictably across four orders of magnitude of compute and reached Pareto-optimal loss versus compute among comparable open base models. The architecture uses DeepSeek MoE-style auxiliary-loss-free load balancing, RMSNorm, and SwiGLU experts. A two-stage midtraining phase stretched the context to 1M tokens, 69 hours at 32K and 24.8 hours at 512K, all on correctness-anchored data, which the paper says did not degrade general capabilities.
Post-training: two teachers, one student
Beam splits competence and compliance into two experts trained separately. The agentic expert goes through SFT then RL, on tasks spanning software engineering, terminal use, mathematics, competitive coding, web search, and professional tool use. The alignment teacher handles safety, calibration, and conversational quality. Multi-teacher on-policy distillation (MOPD) then distills both teachers into a single student. The team even names a merge failure mode, split-braining: the worry that the merged student follows whichever teacher the prompt distribution resembles, instead of genuinely combining both skill sets. The infrastructure numbers are the headline: the high-compute RL configuration reached 1,920 trainer GPUs plus 8,576 inference GPUs, with the inference-to-training ratio drifting between 3.9:1 and 5.4:1. The broader campaign, some 10.5K GB300 GPUs in total including derisking runs and MOPD, exceeded 100M rollouts. The paper calls it one of the largest-scale RL runs conducted by any open lab. The sandbox fleet alone spans more than 20 clusters across two clouds and four regions — environment supply is itself infrastructure for agentic RL.
Two honest observations inside the RL run
First, cross-domain transfer: BrowseComp rose from 50.1 at 27% of the run to 57.9 at 83%, while the task mix over that stretch contained no search or browsing tasks at all. Second, length efficiency is learned, not assumed: the reward includes a controllable length penalty that rewards successful solutions while penalizing output tokens beyond a per-effort budget. Reward hacking gets its correctness score set to zero while keeping its length cost, so cheating is never cheap. The paper also reports no sign of a plateau on its evaluation suite as RL compute kept increasing.
Scores and efficiency
All numbers self-reported. Terminal-Bench v2.1 at 80.1, against Kimi K3 at 88.3, Qwen 3.8 Max at 86.6, and GLM 5.2 at 81.0 in the same table; SWE-Bench Pro v1 at 66.1, SWE-bench Verified at 81.6, AIME 2026 at 97.7, and HLE without tools at 36.0, where the same table lists Kimi K3 at 46.9. Comparisons use each model's recommended sampling settings at the highest available reasoning effort, a 65,536-token output budget per call, and a 256K context window for all models. The serving story is the sales pitch: FP8 and NVFP4 quantization plus DFlash 2, a 1.93B-parameter speculative decoding drafter, lift decode throughput by 1.9x to 3.0x. In BF16, four GPUs hold the weights but leave KV cache room for only 213K tokens; quantized, the same four GPUs serve the full 1M-token context. On the Iris safety suite, Beam refuses or pivots on 88.9% of harmful requests while complying with 81.5% of harmless ones, which the paper calls the highest refusal rate on the Pareto frontier among open-weight models. Under its Open Safety Framework, all four systemic-risk categories are rated Limited.
The cold water
Every number comes from an unreviewed self-reported technical report, with no third-party replication. The paper itself concedes that on raw capability, frontier open models like Kimi K3 remain ahead, and that token efficiency and 3-4x lower inference compute are the actual selling points. The largest-RL-run and highest-refusal claims are its own framing. And the report offers no verifiable download link for the weights, so the open-weight status rests, for now, on the team's own statement.
The number worth remembering is 8,576 versus 1,920: in the RL era, the center of gravity of compute has moved from gradient updates to rollout inference. The open-source bar is shifting from whose base model is stronger to whose RL factory is bigger. Next time a new open model lands, ask one question first: how many of your GPUs were running inference?