Beam is a new open-weight model built for coding, reasoning, and agentic workloads. It is a sparse Mixture-of-Experts architecture with 501 billion total parameters, 23 billion of which activate per token, and it pairs a large pretraining run with reinforcement learning at frontier scale to deliver competitive coding and tool-use performance with lower inference compute than peers.
Weights, a technical report, a model card, and developer artifacts are slated for release later this month after final red-teaming and evaluations, with early access available through signup.
What Beam is and how it was trained
Beam was pretrained on 23.8 trillion curated tokens drawn from the web and proprietary licensed datasets, with the team reporting that this stage matches or outperforms available similar-sized open base models. Training then moved to high-compute reinforcement learning, which the developers call a central scaling axis for the model.
The RL run deployed 10,500 NVIDIA GB300 GPUs for four weeks, generating more than 100 million rollouts with a maximum context length of 256K tokens and using roughly 1.3 billion sandboxes for training and grading. The team describes this as one of the largest-scale RL runs conducted by an open lab to date.
Benchmark performance across coding, reasoning, and tool use
Beam was evaluated against a range of open models including Inkling, Nemotron 3 Ultra, GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8 Max, DeepSeek V4.1 Flash, and DeepSWE v1.1 across coding, reasoning, tool-calling, and general capability benchmarks.
On agentic coding and terminal tasks, Beam scored 44.4 on DeepSWE v1.1, 77.2 on SWE Bench Pro v2-Hard, 65.5 on SWE Bench Pro v1, 80.1 on Terminal Bench v2.1, 34.6 on SWE Atlas Codebase QnA, 78.0 on SWEBench Multilingual, and 80.9 on SWEBench Verified. On reasoning benchmarks it reached 97.8 on AIME 2026, 36.2 on HLE without tools, 49.7 on SciCode, 16.3 on CriPT AA, and 90.5 on GPQA Diamond. In tool-calling and search it posted 37.0 on AutomationBench public, 78.7 on MCP Atlas, 38.0 on tau3 banking, 77.4 on BrowseComp with context management, and 80.1 on DeepSearchQA with context management. General-capability scores included 79.3 on AA-LCR, 65.5 on LongBench v2, 79.7 on IFBench, and 13.0 on AA Omniscience index (public split only).
The developers position Beam as advancing the Western open-weight frontier and being competitive with larger open models like GLM 5.2, while approaching Qwen 3.8 Max on coding and agentic tasks. Where frontier open models like Kimi K3 remain ahead on raw capability, Beam’s reported edge is efficiency at inference time, matching GLM 5.2 on advanced reasoning benchmarks while using 3 to 4 times less inference compute.
How inference efficiency was measured
Compute comparisons used data from Artificial Analysis and DataCurve, estimating generation forward-pass compute as FLOPs approximately equal to 2 times active parameter count times mean generated tokens per attempt, with each multiply-add counted as two operations. Generated tokens include both reasoning and the final answer, and for mixture-of-experts models the parameters activated per token were used rather than total model size. The estimates exclude prompt prefill, context-dependent attention operations, and serving overhead, so they represent an approximate compute comparison rather than measured inference cost.
Reinforcement learning at scale
Scaling RL on Beam required building the algorithms, training environments, and infrastructure to sustain high-compute runs without instability. The team trained with asynchronous policy gradients and developed new algorithms to maintain stable learning while systematically reducing training-inference mismatch. Even when training with one-day staleness, 107 weight versions behind the current policy, numerics remained stable.
Across the evaluation suite, capabilities continued to improve as RL compute increased, with no sign of a plateau. For comparison, Inkling was trained on 30 million rollouts and MiMo on 753,000. On Terminal-Bench 2.1, HLE, and DeepSWE, scores rose as cumulative RL rollouts grew during training of Beam’s reasoning expert, which alone accounted for 80 million of the 100 million plus rollouts generated across the full campaign.
Controllable reasoning effort
Beam was trained with a controllable length penalty that rewards successful solutions while discouraging unnecessary tokens. Early in RL, performance improved even as completion lengths fell as the model learned to solve tasks more effectively with less reasoning. Later, as Beam developed stronger agentic capabilities, completion lengths grew again, with those additional tokens supporting further performance gains. Throughout training, RL improved the trade-off between capability and token usage, and users can control this trade-off through a reasoning effort parameter. Lower settings favor shorter responses, while higher settings allow longer reasoning to improve performance on demanding tasks.
How behavior generalizes
During a phase of training on reasoning, software engineering, and terminal tasks, Beam showed consistent gains in browsing despite the absence of browsing tasks from the RL mixture, suggesting it was learning broader agentic capabilities that generalize across domains. When given web access, Beam organically learned to search for and query other large language models and to use OCR APIs to read documents.
Demo examples include a 3D astronaut free-fall game built in p5.js with dodge-or-destroy mechanics, a 180 by 90 longitude-latitude grid puzzle where Beam reached 95.5 percent coverage between two reference models, a live-updating NYC subway map built from public MTA data, and a fine-tuning notebook for a small Gemma-4 model on a Text2SQL task that improved held-out test accuracy by 66.5 percent. Although Beam is text-only, it can work with information from other modalities when represented as text.
Environments and infrastructure
Frontier-scale RL required a large volume of difficult, high-quality tasks. The team built a pool of nearly one million environments, primarily through synthetic data pipelines supplemented by proprietary vendor data and open-source sources, covering software engineering, terminal use, competitive coding, STEM, web search, tool use, and general knowledge work. Tasks were iteratively filtered for difficulty and quality, then tested through RL to inform the next round of sourcing and filtering.
The training platform runs fully asynchronously, sustaining an average of 110,000 concurrent rollouts. Compute allocation between inference and training was adjusted as the workload evolved, with inference-to-training GPU ratios ranging from 3.9:1 to 5.4:1, and the trainer was resized across five GPU mesh configurations within the same training lineage without losing training state. New weights reached the inference fleet in a median of about 12 seconds using a hierarchical distribution that transfers weights across racks over RoCE and shares them locally over NVLink, reducing cross-rack traffic by 75 percent and making fleet-wide adoption 2.2 times faster. The system handled 71 inference incidents without terminating the training job, with recovery in a median of eight minutes and lost capacity at 0.02 percent of elapsed serving GPU-minutes. The platform supported up to 170,000 environments at peak scale.
What the underlying research shows
Beam’s mixture-of-experts design sits in a lineage of open models that have explored large sparse architectures. The DeepSeek-V3 technical report, published on arXiv as 2412.19437, describes a 671 billion total parameter MoE with 37 billion activated per token, trained on 14.8 trillion tokens using Multi-head Latent Attention and an auxiliary-loss-free load balancing strategy, with full training completed in 2.788 million H800 GPU hours. DeepSeek-V3 reported performance comparable to leading closed-source models at the time, and its training was described as remarkably stable with no irrecoverable loss spikes or rollbacks.
On the safety side, the deliberative alignment paper, published on arXiv as 2412.16339, introduced a paradigm that directly teaches models safety specifications and trains them to explicitly recall and reason over those specifications before answering. The approach was used to align a series of reasoning models, and the paper reported higher robustness to jailbreaks, lower overrefusal rates, and improved out-of-distribution generalization compared with prior alignment methods.
What to watch next
Three things will define Beam’s real-world standing after release. First, the weight download and technical report will let independent labs reproduce the benchmark numbers and probe the inference-efficiency claims on their own hardware. Second, the RL infrastructure details, from the 110,000 concurrent rollouts to the 12-second weight propagation, will inform how other open labs budget their next training runs. Third, the reasoning effort parameter will determine whether developers can actually dial the capability-versus-token trade-off to match their cost targets in production coding and agentic deployments.
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.
How large was Beam’s reinforcement learning run?
The RL run used 10,500 NVIDIA GB300 GPUs for four weeks, generated more than 100 million rollouts with a 256K token context length, and used approximately 1.3 billion sandboxes across nearly one million training environments.
How does Beam compare to other open models on coding benchmarks?
On agentic coding and terminal tasks Beam scored 80.9 on SWEBench Verified, 80.1 on Terminal Bench v2.1, 77.2 on SWE Bench Pro v2-Hard, 65.5 on SWE Bench Pro v1, and 44.4 on DeepSWE v1.1, with the developers reporting competitive performance against larger open models like GLM 5.2 and approaching Qwen 3.8 Max.
Related coverage
- Qwen3.6-27B: Open-Source 27B Model Beats 397B MoE on Coding Benchmarks
- Cisco Antares: Open-Weight AI Models for Vulnerability Localization
- Poolside Laguna S 2.1 open-weight coding model
Want help with this in Myrtle Beach or Long Island? Talk to us.
NewSunSEO
Grow in Google and AI Search
AI-SEO, local search, paid ads and websites from one team, with offices on Long Island and in Myrtle Beach.
- AI-SEO
- Local AI-SEO
- AI Consulting
- PPC
- Web Design
- Web Development
- Long Island Office(631) 320-1700
- Myrtle Beach Office(843) 492-7131
Request a Free Proposal
Tell us about your business and where you want to grow.
We will reply by email to talk through your goals and what a proposal would include.
We use your details only to reply to you and send what you asked for. See our Privacy Policy.
