Introducing Beam: A 501B Open-Weight Model for Coding and Agentic Workloads

Readers running coding or agentic workloads get a new open-weight option: Beam, a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active per token. Beam pairs coding and agentic focus with frontier-level inference efficiency, and its weights, technical report, and model card are scheduled to release later this month.

What Beam is and what it targets

Beam is built for coding, reasoning, and agentic workloads. The model advances the Western open-weight frontier and is competitive with larger open models on coding and agentic tasks while approaching Qwen 3.8-Max in those areas. Where frontier open models like Kimi K3 remain ahead on raw capability, Beam’s stated advantage is efficiency at inference time: comparable reasoning scores to GLM-5.2 using 3 to 4 times less inference compute, and large efficiency gains against 2T+ parameter models that need significantly more compute per token.

That efficiency profile is designed to deliver more intelligence per token at lower cost, positioning Beam as a workhorse model for enterprise coding and agentic deployments.

Benchmark results across coding, reasoning, and tool use

Beam was evaluated against Inkling, Nemotron 3 Ultra, GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8 Max, DeepSeek V4.1 Flash, and DeepSWE v1.1. Selected scores (NR marks scores that have not been reported):

• DeepSWE v1.1: Beam 44.4, GLM 5.2 44.0, GLM 5.3 61.0, Kimi K3 68.0, Qwen 3.8 Max 51.0, DeepSeek V4.1 Flash 74.2.
• SWE Bench Pro v2-Hard: Beam 77.2, Inkling 56.9, GLM 5.3 84.3, Kimi K3 88.2.
• SWE Bench Pro v1: Beam 65.5, Inkling 54.3, Nemotron 3 Ultra 46.4, GLM 5.2 62.1, Qwen 3.8 Max 67.7.
• Terminal Bench v2.1: Beam 80.1, Inkling 63.8, Nemotron 3 Ultra 56.4, GLM 5.2 81.0, GLM 5.3 88.2, Kimi K3 88.3, Qwen 3.8 Max 86.6, DeepSeek V4.1 Flash 90.6.
• SWEBench Verified: Beam 80.9, Inkling 77.6, Nemotron 3 Ultra 70.7.
• AIME 2026: Beam 97.8, Inkling 97.1, GLM 5.2 99.2.
• HLE no tools: Beam 36.2, Inkling 29.7, Nemotron 3 Ultra 26.7, GLM 5.2 40.5, GLM 5.3 42.3, Kimi K3 46.9, Qwen 3.8 Max 43.6, DeepSeek V4.1 Flash 39.1.
• GPQA Diamond: Beam 90.5, Inkling 87.2, Nemotron 3 Ultra 87.0, GLM 5.2 91.2, GLM 5.3 91.7, Kimi K3 93.5, Qwen 3.8 Max 92.6, DeepSeek V4.1 Flash 90.9.
• AutomationBench public: Beam 37.0, GLM 5.2 26.2, GLM 5.3 48.2, Kimi K3 46.7, Qwen 3.8 Max 39.8, DeepSeek V4.1 Flash 54.8.
• MCP Atlas: Beam 78.7, Inkling 76.0, Nemotron 3 Ultra 63.1, GLM 5.2 77.8, GLM 5.3 84.2, Kimi K3 82.3, Qwen 3.8 Max 84.5.
• tau3 banking: Beam 38.0, Inkling 25.0, Nemotron 3 Ultra 22.6, GLM 5.2 37.1, Kimi K3 37.1, Qwen 3.8 Max 55.2.
• BrowseComp with context management: Beam 77.4, Inkling 77.1, Nemotron 3 Ultra 44.4, Kimi K3 91.2.
• DeepSearchQA with context management: Beam 80.1, Kimi K3 95.0.
• AA-LCR: Beam 79.3, Inkling 77.3, Nemotron 3 Ultra 79.3, GLM 5.2 78.3, GLM 5.3 79.7, Kimi K3 88.7, Qwen 3.8 Max 80.3, DeepSeek V4.1 Flash 84.0.
• LongBench v2: Beam 65.5, Nemotron 3 Ultra 61.9, GLM 5.2 64.0, Qwen 3.8 Max 66.3.

External benchmark sources include Artificial Analysis and DataCurve. For the inference-compute comparison, generation forward-pass compute is estimated as FLOPs roughly 2 times active parameter count times mean generated tokens per attempt, with each multiply-add counted as two operations. Generated tokens include reasoning and the final answer. For MoE models, active parameters per token are used rather than total model size. Prompt prefill, context-dependent attention, and serving overhead are excluded, so the figures approximate compute rather than measured inference cost.

Pretraining a foundation for reasoning

Beam was pretrained on 23.8 trillion diverse, curated, high-quality tokens drawn from the web and proprietary licensed datasets, matching or outperforming available similar-sized open base models. In parallel, the algorithms, training environments, and infrastructure for high-compute reinforcement learning were developed.

High-compute reinforcement learning at scale

High-compute reinforcement learning was treated as a central scaling axis. The training run deployed 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 approximately 1.3 billion sandboxes. To sustain the run, one million high-quality coding, agentic, and STEM environments were sourced, and capabilities continued to improve as RL compute increased, with no sign of a plateau across the evaluation suite.

During RL on reasoning, software engineering, and terminal tasks, consistent gains in browsing appeared despite the absence of browsing tasks from the RL mixture, indicating broader agentic capabilities that generalize across domains. With web access, the model organically learned to search for and query other large language models and to use OCR APIs to read documents.

Asynchronous training and stable numerics

Beam was trained with asynchronous policy gradients. At scale, policy staleness is a major source of instability, and numerical mismatch between training and inference engines compounds it. New algorithms maintained stable learning under these conditions and systematically reduced training-inference mismatch. Even with one-day staleness, 107 weight versions behind the current policy, the numerics remained stable.

For comparison, Inkling was trained on 30 million rollouts and MiMo on 753,000. The Terminal-Bench 2.1, HLE, and DeepSWE scores plotted against cumulative RL rollouts show continued gains over 80 million of the 100M+ rollouts on Beam’s reasoning expert.

Learning to reason efficiently

Beam was trained with a controllable length penalty that rewards successful solutions while discouraging unnecessary tokens. Early in RL, performance improved as completion lengths fell: the model learned to solve tasks with less reasoning. Later, as agentic capabilities strengthened, completion lengths grew again, with the additional tokens supporting further performance gains. A reasoning effort parameter lets users tune this tradeoff: lower settings favor shorter responses, higher settings allow longer reasoning for demanding tasks.

Frontier-scale RL infrastructure

Frontier agentic RL requires generating rollouts, executing tools, evaluating outcomes, and updating the model at scale. The asynchronous platform sustained an average of 110K concurrent rollouts during training, with inference-to-training GPU ratios adjusted from 3.9:1 to 5.4:1 and the trainer resized across five GPU mesh configurations within the same training lineage without losing state. New weights reached the inference fleet in a median of approximately 12 seconds via hierarchical distribution over RoCE and NVLink, cutting cross-rack traffic by 75% and making fleet-wide adoption 2.2 times faster. Across the run, 71 inference incidents were handled without terminating the training job, with a median recovery of eight minutes and lost capacity at 0.02% of elapsed serving GPU-minutes.

Capability demos across novel tasks

Although Beam is text-only, it can work with information from other modalities when represented as text. A recreated X geography puzzle asked Beam to build a fixed 180 by 90 grid spanning longitudes -179 to 179 and latitudes -89 to 89, with 16,200 points; Beam reached 95.5% coverage, between Opus 5 at 92.5% and Fable 5 at 97.8%. Beam was also prompted to build a live updating NYC subway map using publicly available MTA data, searching documentation, checking authentication, finding subway and map geometries, and building the right frontend and backend with a server that keeps running for live updates.

In an out-of-distribution task, Beam was plugged into OpenCode, asked to consult Unsloth, and produce a fine-tuning notebook for the smallest Gemma-4 model on a Text2SQL task. It researched the model cards, Unsloth documentation, and READMEs, and implemented the fine-tuning logic, increasing Gemma-4’s accuracy on its held-out test set by 66.5%.

Safety, alignment, and the path ahead

Beam is undergoing final red-teaming and evaluations. Sign-ups are open for early access, with weights, technical report, model card, and developer artifacts to follow later this month. A separate research paper on safety methodology, Deliberative Alignment: Reasoning Enables Safer Language Models (arXiv:2412.16339), describes a paradigm that directly teaches models safety specifications and trains them to explicitly recall and reason over the specifications before answering, achieving precise adherence to safety policies without human-written chain-of-thought, increasing robustness to jailbreaks while decreasing overrefusal rates, and improving out-of-distribution generalization.

MoE training methodology has matured in the open over the past year. The DeepSeek-V3 Technical Report (arXiv:2412.19437) documents a 671B total parameter MoE with 37B activated per token, pretrained on 14.8 trillion tokens using Multi-head Latent Attention and an auxiliary-loss-free load balancing strategy, with full training completed in 2.788M H800 GPU hours, no irrecoverable loss spikes, and no rollbacks. That work validated the kind of large-scale MoE training and inference efficiency that underpins Beam’s design.

FAQ

What is Beam?

Beam is an open-weight sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active per token, built for coding, reasoning, and agentic workloads. Weights, technical report, model card, and developer artifacts are scheduled to release later this month.

How is Beam different from other open-weight models?

Beam focuses on coding and agentic capability while targeting frontier-level inference efficiency. On advanced reasoning benchmarks, it reaches scores comparable to GLM-5.2 while using 3 to 4 times less inference compute, and it shows large efficiency gains against 2T+ parameter family models.

How was Beam trained?

Beam was pretrained on 23.8 trillion tokens and then trained with high-compute reinforcement learning using 10.5K NVIDIA GB300 GPUs for four weeks, generating more than 100 million rollouts with a maximum context length of 256K tokens and approximately 1.3 billion sandboxes, sourced from a pool of one million coding, agentic, and STEM environments.