Reflection AI Unveils Beam, a 501B Open-Weight Reasoning Model
Reflection AI has unveiled Beam, its first frontier open-weight AI model. TechCrunch The company detailed the system on Oct. 5, 2026.
Beam is a text-only mixture-of-experts model trained with high-compute reinforcement learning for reasoning, coding and agentic tasks. Reflection The text-only constraint is notable. It narrows the optimization surface to tokens, tool traces and code execution, rather than spreading capacity across vision or audio.
The parameter structure is 501 billion total parameters with 23 billion active parameters. Pre-training covered 23.8 trillion tokens, and the model supports a 1 million token context window. TechCrunch That is a sparse activation ratio. Only a small fraction of weights engage per forward pass.
Reflection claims Beam scores on par with Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. GLM-5.2 is the relevant comparator here, with roughly 744 billion total parameters and 40 billion active parameters. Reflection also claims Beam outperforms leading Western open models on advanced reasoning benchmarks. The claims are framed around inference efficiency, not just raw score. For deployment teams, that distinction matters more than leaderboard position.
The training methodology centers on reinforcement learning scale. Reflection built a pool of nearly one million environments for Beam's reinforcement learning training. Those environments were created primarily through synthetic data pipelines, supplemented by proprietary vendor data and open-source sources. Reflection
Across the full reinforcement learning campaign, Reflection generated over 100 million rollouts. Eighty million of those rollouts were used to train Beam's reasoning expert alone. For comparison, Reflection states Inkling was trained on 30 million rollouts and MiMo on 753,000 rollouts. The order-of-magnitude jump is explicit. It points to environment diversity and rollout volume as the lever, rather than pre-training token count alone.
Reflection said it used Artificial Analysis and DataCurve as sources for other models' evaluations in its inference-efficiency comparisons. That sourcing detail is useful for practitioners who track third-party inference pricing and throughput. It does not resolve benchmark comparability on its own, but it makes the efficiency claim auditable against public evaluation infrastructure.
For context on the comparator, GLM-5.2 is available on Z.ai, and its model weights are publicly available on HuggingFace and ModelScope. Z.ai That availability is why the Reflection comparison lands with Western labs and enterprise builders. Both systems are accessible as open weights, which allows direct testing on internal reasoning and coding suites.
Reflection was founded in 2024 by two former Google DeepMind researchers. It has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners, per PitchBook. The funding scale explains how a two-year-old lab can sustain a 23.8 trillion token pre-train followed by a 100 million rollout reinforcement learning phase.
Compute access is secured separately. Reflection signed deals collectively worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia GB300 chips through 2029. That timeline extends well beyond Beam. It covers future training runs and inference capacity, and it ties accelerator supply to nontraditional cloud partners.
The broader context here is a shift in where open-model competition happens. Pre-training scale once dominated discussion. Active parameters and inference cost per reasoning task now shape adoption. A 23 billion active MoE that matches a 40 billion active MoE at lower compute would change serving economics for coding assistants, software agents and long-context workflows, where margin is set by tokens generated per dollar and latency under load.
In my view, the number worth watching is not 501 billion. It is nearly one million environments. Worth flagging: synthetic environment generation at that scale introduces its own failure modes around distribution skew and reward hacking, which only emerge under extended agentic rollouts. If Reflection has managed to keep environment quality high while pushing to 100 million rollouts, that pipeline is the durable asset. It would transfer to future models more directly than any single checkpoint, and it would give enterprise teams an open-weight option tuned for verifiable, multi-step work rather than single-turn response.
