Beam's weights are still pending
Reflection's Beam is a 501B-parameter sparse open-weight model, but its weights and technical artifacts are still pending release.
Reflection says Beam has 501 billion total parameters, but only 23 billion active per token. That is the concrete shape of the model behind the announcement: a sparse Mixture-of-Experts system aimed at coding, reasoning, and agentic workloads.
The weights have not been released in the source material. Reflection describes Beam as its first open-weight model, says it is in final red-teaming and evaluations, and says the weights, technical report, model card, and developer artifacts will be released later this month under Apache 2.0. For now, the described access path is early access via signup.
The interesting part is less the branding and more the training system around it. Beam was pretrained on 23.8 trillion tokens from web data, public sources, and proprietary licensed datasets. Reflection says it trained on almost all publicly accessible and unrestrictively licensed code and code documentation on the web, then filtered heavily: about 95% of raw Internet tokens were eliminated through parsing, deduplication, and curation. The company also says conventional filtering would have missed roughly 1.8 trillion high-quality tokens it retained, including 87% of its curated web-code tokens.
That matters because this is a coding and agentic model, not just a chat model with a benchmark table attached. Reflection says Beam was trained with a particular focus on coding and agentic performance, and that its advantage is inference efficiency compared with larger open models. The claim is that Beam approaches stronger systems on coding and agentic tasks while using less inference compute on reasoning benchmarks. The source gives the most specific comparison as scores comparable to GLM-5.2 on advanced reasoning benchmarks while using 3 to 4 times less inference compute.
The reinforcement learning run is the other large number in the announcement. Reflection says Beam used 10.5K NVIDIA GB300 GPUs over four weeks to generate more than 100 million rollouts. Training and grading used about 1.3 billion sandboxes. The platform sustained an average of 110K concurrent rollouts and supported up to 170K concurrent sandboxes during the run.
For engineers running agentic systems, the sandbox count is the part to stare at for a while. Agentic RL at this scale is mostly an infrastructure problem wearing a model-training jacket. The system has to generate rollouts, execute tools, grade outcomes, move weights through an inference fleet, and keep training stable while experience arrives late. Reflection says its setup was fully asynchronous: agents generated rollouts while the trainer learned and published new model versions. Each token was tagged with the model version that produced it so training could account for policy staleness.
That is a useful implementation detail. Once rollouts run long enough, the data stream stops being cleanly tied to a single current policy. Earlier tokens can come from older checkpoints while later training has already moved on. Reflection says it developed algorithms to keep asynchronous RL stable even when learning from interactions generated more than a day earlier. It also calls out numerical mismatch between training and inference engines as a compounding problem.
The weight distribution system is another operational signal. During Beam training, new weights reached the inference fleet in a median of about 12 seconds. Reflection says hierarchical distribution moved weights across racks over RoCE, then shared them locally over NVLink. Compared with every replica pulling weights directly, that reduced cross-rack traffic by 75% and made fleet-wide adoption of new weights 2.2 times faster.
That kind of detail is useful because it turns “frontier RL infrastructure” into something less mystical. Large agentic training runs need boring properties: fast rollout generation, controlled checkpoint propagation, replayable records, failure handling, and enough observability to know whether rewards are being gamed. Reflection says 71 inference incidents were handled without terminating the training job, with inference capacity recovering in a median of eight minutes. Lost capacity accounted for 0.02% of elapsed serving GPU-minutes.
Beam also has an explicit length-control story. Reflection says it trained the model with a controllable length penalty that rewards successful solutions while discouraging unnecessary tokens. Early in RL, performance improved while completion lengths fell. Later, completion lengths grew again as agentic capabilities improved, with the added tokens supporting more performance. Users get a reasoning effort parameter: lower settings favor shorter responses, higher settings allow longer reasoning on harder tasks.
That is likely to matter in production more than another small benchmark delta. Agentic coding systems are often constrained by latency, tool-call cost, and context budget before they are constrained by raw model capability. A model that can expose a usable effort dial gives operators a clearer way to trade answer quality against spend. The source does not give the API shape, but the control itself is part of the design.
The pretraining infrastructure was separate and also large. Reflection says Beam was pretrained end-to-end in under four weeks on 6,144 NVIDIA GB300 NVL72 GPUs. The company built most of the infrastructure stack in-house, including a topology-aware Kubernetes scheduler, a node lifecycle system with health monitoring and alerts, and a silent data corruption detection system capable of semi-autonomous rewinds and restarts. The run had nine semi-automatic rewinds, attributed either to non-deterministic gradient norm spikes or suspected silent data corruption. Goodput reached 92.3% toward the end.
The safety and alignment section is more process than result. Reflection says it trained a second model from the pretrained checkpoint using a separate SFT and RL pipeline for safety and alignment principles, then merged capabilities from a large-scale RL teacher and a dedicated safety teacher via multi-teacher on-policy distillation. It says it will publish safety evaluation results in the technical report and open-source internal safety evaluations.
For teams planning to run Beam when the weights arrive, the immediate implication is simple: treat this as a pending open-weight release, not an artifact already available for local deployment. The source says the release will include weights, documentation, and the full stack for running, evaluating, and fine-tuning the model. Until those artifacts exist, the engineering work is mostly preparation: capacity planning for a 501B-parameter sparse model, evaluation harnesses for coding and agentic workloads, and deployment assumptions around reasoning-effort controls rather than fixed completion behavior.
The announcement supports interest, not procurement-grade certainty. The technical report, model card, developer artifacts, safety evaluations, and the weights themselves are still the pieces that will determine how Beam fits into real systems.
Keep Reading
AIStanford teaches LLMs by making you build one
Stanford CS336 teaches language models by having students build the training stack end to end, from tokenizers to data and alignment.
AIWhat distillation leaves behind
Y Combinator's Garry Tan opposes regulating AI distillation and wants US open-weight labs to distill American frontier models through the front door.
AIDistillation makes fast followers, never frontier leaders
Garry Tan wants US open-weight labs to distill frontier models legally. Distillation copies a teacher, so it follows the frontier but can't lead it.
Latest on the Wire
Full wire →- AI Coding Agents Boost Output But Code Review Bottlenecks IncreaseHacker News
- AI Overviews Cut Organic Traffic by 40%Hacker News
- Apache Patches Critical Roller RCE VulnerabilityBleepingComputer
- Complete Web Browser Built from Scratch in PythonHacker News
New signal daily · RSS
Stay in the loop
New writing delivered when it's ready. No schedule, no spam.