Reflection Previews Beam, a 501B-Parameter Coding Model Built on 100 Million RL Rollouts

Reflection AI introduced Beam on October 5, 2026, calling it the company’s first open-weight model. Beam is a sparse mixture-of-experts model with 501 billion total parameters, 23 billion of which are active per token. Reflection says it is built for coding, reasoning, and agentic workloads. The weights are not public yet. Reflection says Beam is “undergoing final red-teaming and evaluations,” that early access is through a waitlist, and that it “will release the weights, technical report, model card, and developer artifacts later this month” under the Apache 2.0 license, along with tooling for running, evaluating, and fine-tuning the model.

Where the benchmarks place it

Reflection does not claim Beam leads the open field. Its post says Beam is “competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks,” and it concedes that frontier open models like Kimi K3 “remain ahead on raw capability.” Its pitch is efficiency at inference time.

The company’s tables show that trade-off. Beam scores 80.9 on SWE-bench Verified and 78.0 on SWE-bench Multilingual. On SWE Bench Pro v1 it scores 65.5, against 62.1 for GLM 5.2 and 67.7 for Qwen 3.8 Max. On Terminal Bench v2.1 it scores 80.1, behind GLM 5.2 (81.0), Kimi K3 (88.3), and DeepSeek V4.1 Flash (90.6). In reasoning, Beam reports 97.8 on AIME 2026 and 90.5 on GPQA Diamond. On Humanity’s Last Exam without tools it scores 36.2, under GLM 5.2’s 40.5.

The efficiency claim is that Beam matches GLM-5.2 on advanced reasoning “while using 3–4× less inference compute.” Reflection explains how it got that number. It estimates generation compute as two times the active parameter count times the mean number of generated tokens, using comparison data from Artificial Analysis and DataCurve. It notes that the estimate leaves out prompt prefill, context-dependent attention, and serving overhead, so it is “an approximate compute comparison rather than measured inference cost.” All of these scores are Reflection’s own. None has been independently reproduced, because the weights are not out.

A reinforcement learning run at fleet scale

Most of the post covers training. Reflection says reinforcement learning ran on 10.5K NVIDIA GB300 GPUs for four weeks and generated more than 100 million rollouts, with a maximum context length of 256K tokens. Training and grading used about 1.3 billion sandboxes, drawn from a pool of nearly one million coding, agentic, and STEM environments that were built mostly through synthetic pipelines. For comparison, the company says Inkling was trained on 30M rollouts and MiMo on 753K.

The training was fully asynchronous. Each token is tagged with the model version that produced it, so the trainer can account for “policy staleness.” Reflection says learning stayed stable even with one-day staleness, 107 weight versions behind the current policy. Across the run it averaged 110K concurrent rollouts and peaked at 170K concurrent sandboxes. New weights reached the inference fleet in a median of about 12 seconds, and the job survived 71 inference incidents without being terminated.

A controllable length penalty was used to teach Beam to reason with fewer tokens. Users get a matching reasoning-effort setting that trades response length against accuracy. The company also reports a transfer effect: after training on reasoning, software engineering, and terminal tasks, Beam got better at browsing even though browsing tasks were not in that RL mix.

The base model underneath

Beam’s base was pretrained on 23.8 trillion tokens from the web, public sources, and proprietary licensed datasets. The run took under four weeks on 6,144 NVIDIA GB300 NVL72 GPUs. Reflection reports nine semi-automatic rewinds during the run and goodput of 92.3% toward the end. Its load balancing kept expert use nearly uniform: averaged across MoE layers, the busiest expert’s load was 1.04× at the end of pretraining. Midtraining extended Beam’s effective context length to 1M tokens. Beam is text-only, though the company says it can work with other modalities when they are represented as text.

Safety work ran as a separate pipeline. Reflection trained a second model from the pretrained checkpoint for safety and alignment, then merged that teacher with the large-scale RL teacher through multi-teacher on-policy distillation. It says safety evaluation results will appear in the technical report, and that the safety evaluations it developed internally will be open-sourced.

What is still missing

As of the announcement, there is no model card, no license file, no published checkpoint, and no third-party evaluation. The architecture details shared so far are partial: interleaved local and global attention, fine-grained routed experts, and 52 layers. Builders comparing Beam with GLM, Kimi, or DeepSeek models that are already downloadable will have to wait for the October release to check the numbers. Reflection describes Beam as “the first model in a series” and says it is already training what comes next.