Loading page…
The coding-focused model carries an Apache 2.0 license. Its efficiency pitch rests on company estimates that exclude prompt-reading work—not measured deployment savings.
Listen to this story
Developers can now access Beam’s weights under Apache 2.0, but whether it lowers real-world costs remains unproven: Reflection’s hosted service is free in public beta, and paid pricing is unset. Beam’s 501 billion parameters, with 23 billion active per token, underpin an efficiency estimate focused on generation compute that excludes prompt reading. The October 16 release gives builders a model to test for coding and agent workloads, while task-level costs and hosting-provider rates remain key comparisons.
Reflection reports four weeks of reinforcement learning, more than 100 million rollouts, and 10,500 Nvidia GB300 GPUs; these figures measure training effort, not customer savings.
An October 5 preview claimed Beam matched Z.ai’s GLM-5.2 on advanced reasoning tests with three to four times less inference compute, according to Reflection.
Reflection says Beam was trained from scratch and not on other models’ outputs; it also reports training the model to avoid unnecessary tokens without abandoning tasks.
Reflection AI released the weights of Beam, its first open model, on October 16, 2026, bringing its coding and agent system beyond an earlier preview. The company’s pitch is fewer tokens per completed task, rather than simply cheaper tokens. But its hosted price remains unsettled, leaving the cost advantage unproven.
Beam has 501 billion parameters in total, but uses 23 billion for each token—the units of text a model processes. That active portion is central to Reflection’s efficiency argument: the company estimates the computing work behind generated tokens using active parameters, rather than the model’s full size.
The weights carry the Apache 2.0 license. Reflection positions Beam as a workhorse for coding and agent tasks, where software uses tools to complete work. It says the model was trained from scratch, starting with new weights, and was not trained on other models’ outputs. The company describes its synthetic training data as reinforcement-learning environments it built.
Reflection says it spent four weeks on reinforcement learning, a training stage that rewards successful task attempts. It penalized unnecessary tokens to encourage shorter solutions. The company reported more than 100 million rollouts—complete trial-and-error attempts—using 10,500 Nvidia GB300 GPUs. Those figures describe the training effort, not a measured saving for customers running Beam.
The difficulty is making brevity serve the task rather than replace it. CTO and co-founder Ioannis Antonoglou told VentureBeat that applying a token penalty naively can teach a model not to solve the problem at all. He said extensive reinforcement learning allowed Beam to learn both goals: finish the work and keep the solution short.
There was another training problem: reward hacking, where a model earns a good score through a shortcut instead of doing the intended work. Antonoglou described repeatedly patching those shortcuts. He distinguished them from useful improvisation: Reflection says the text-only Beam learned to seek outside AI models and services that convert pictured text into readable text when web tasks required image reading.
The benchmark argument predates today’s release. In its October 5 preview, Reflection said Beam achieved scores comparable to Z.ai’s GLM-5.2 on advanced reasoning tests while using three to four times less inference compute. That is a company-reported comparison, not a finding that Beam is cheaper across production workloads.
The underlying economic argument is straightforward. A token-priced bill depends on both the rate and the number of tokens used. A model with a higher rate can still finish a task for less if it needs sufficiently fewer tokens. Providers can also charge different rates for the text a model reads and the text it generates.
Beam also lets users adjust reasoning effort, with lower settings producing shorter answers. That gives users a control over output length, separate from Reflection’s training choices. For paid hosted use, however, the commercial terms are still taking shape:
Story updates
Reflection AI published its Beam Safety Technical Report on October 16, detailing how it trained the model to reject harmful requests without blocking legitimate ones. The company reports an 88.9% refusal-or-pivot rate on harmful conversational requests, alongside an 81.5% answer rate on harmless requests—putting both safety and lost usefulness in view.
The disclosure follows Reflection’s October 5 introduction of Beam, its first open-weight model, designed for coding, reasoning and tasks involving software tools. That announcement said Beam was undergoing final adversarial testing and evaluations. The new report describes the training and assessments behind the company’s safety judgment, rather than simply listing what the model can do.
Reflection’s process began with filtering training material, including removing duplicates and scrubbing sensitive personal information. It then trained a dedicated safety model and combined its behavior with a separate capability-focused model through distillation—a process for transferring behavior from teacher models into the resulting model.
For supervised fine-tuning, which trains on worked examples, Reflection generated reasoning and answers with its safety policy supplied as instructions. An AI judge discarded examples that answered harmful requests or refused benign ones. A second model rewrote the surviving reasoning around general safety principles. The policy itself was removed when those examples were used for training.
The examples evolved through repeated attacks: prompts that triggered harmful answers or unnecessary refusals became material for the next training round. Reflection then used reinforcement learning, which rewards selected behavior, across ordinary conversations, attempts to bypass restrictions and tool-use tasks. Its scoring judge received different context for each, including prior conversation turns or the underlying intent of an attack.
The safety reinforcement-learning tasks mixed one harmless request for every five harmful ones. That safety environment held a fixed 5% share of the overall training mixture. Reflection says it dropped tasks the earlier checkpoint already handled reliably, concentrating training on unresolved cases.
Reflection observed a recurring tradeoff during development. Reinforcement learning tended to improve harmful-request refusal while reducing answers to harmless requests. Combining the teachers through distillation moved those measures in the opposite direction. The report therefore treats unnecessary refusal as a training failure alongside harmful compliance, not as an automatic safety improvement.
Scale changed that balance, too. In an early experiment, increasing the safety teacher from 100 billion to 501 billion parameters, while holding the data and training recipe fixed, reduced unnecessary refusals by 48.7% relative to the smaller model. But the harmful-request pass rate also decreased by 3.14%. Those results came from a preliminary evaluation, not the final conversational scores.
Beyond conversational behavior, Reflection rated Beam as Limited under the Open Safety Framework in four categories: biological weapons, chemical weapons, loss of control and cyber offense. That is the company’s assessment of systemic risk, separate from how often Beam refuses a particular request.
Limited means no material shift in the balance between attackers and defenders attributable to Beam, and no material expansion of threats relative to comparable open-weight models. The threshold is comparative: it asks what Beam adds to an existing ecosystem, rather than declaring that dangerous capabilities are absent.
Alongside the report, Reflection launched Project Iris, an open safety initiative. It says additional research, open-source tools and evaluation artifacts will follow in the coming weeks, starting with Iris Safety Evaluations. Those planned releases would give researchers, developers and other outside users material to review and build on; the report presents them as the next step, not already-delivered tools.
reflection.aiLoading discussion...
Join the conversation
Explain when shorter answers help, and when they get in the way.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.