A New Shared AI That Reasons for Less

Reflection AI released Beam, its first top-level openly shared AI model, on Oct. 5, 2026. TechCrunch
What Beam is
Beam works with text only. It is a mixture-of-experts model, like a team of specialists where only the right experts step in for each question. It was trained with large-scale reinforcement learning, learning from rewards for correct step-by-step answers, for reasoning, coding and agentic tasks where software uses tools to finish work. Reflection
The text-only limit keeps training focused. It covers tokens, small pieces of text, plus tool traces and code execution. It does not spread effort across vision or audio.
Beam has 501 billion total parameters and 23 billion active parameters. Think of total as the full library, and active as the few books opened for one answer. Pre-training covered 23.8 trillion tokens, and the model can keep 1 million tokens in mind at once. TechCrunch Only a small share of the model works on each answer.
Reflection claims Beam matches Z.ai's GLM-5.2 on hard reasoning tests while using 3-4x less inference compute, the computer power needed to answer users. GLM-5.2 has about 744 billion total parameters and 40 billion active parameters. Reflection also claims Beam beats leading Western open models on hard reasoning tests. The point is lower running cost, not just higher score. For teams that run these models, that cost difference matters more than rank.
GLM-5.2 is available on Z.ai, and its weights are public on HuggingFace and ModelScope. Z.ai Both models are open weights, so others can test them on their own coding and reasoning tasks. That shared access is why the comparison matters to Western labs and company builders.
How it was trained
Reflection built almost one million environments for training. These are like practice tests with answers that can be checked. They were made mostly with synthetic data pipelines, plus proprietary vendor data and open-source sources. Reflection
In total, Reflection ran over 100 million rollouts, practice tries where the model attempts a task and learns. Eighty million rollouts trained the reasoning expert alone. By comparison, Inkling used 30 million rollouts and MiMo used 753,000. The jump points to practice variety and trial volume as the lever, rather than pre-training size alone.
Reflection said it used Artificial Analysis and DataCurve as sources for other models' scores in its cost comparisons. That helps people who follow public prices and speed tests. It does not settle test fairness on its own, but it lets others check the cost claim.
Funding and computing power
Reflection was started in 2024 by two former Google DeepMind researchers. It has raised about $4.7 billion from Nvidia, Sequoia Capital and Lightspeed Venture Partners, per PitchBook. That budget supports a 23.8 trillion token pre-train plus a 100 million rollout training phase at a two-year-old lab.
Computer access was arranged separately. Reflection signed deals worth more than $7 billion with SpaceX and Nebius for Nvidia GB300 chips through 2029. That period goes well past Beam. It covers future training and user traffic, and it links chip supply to less usual cloud partners.
The broader context here is a change in what decides open-model use. Pre-training size once led the talk. Now active size and cost per reasoning task shape adoption. A 23 billion active model that matches a 40 billion active model for less compute would change costs for coding helpers, software agents and long-document work, where profit rests on answers per dollar and speed under load.
In my view, the number to watch is not 501 billion. It is almost one million environments. Worth flagging, making practice tests by machine at that scale can bring errors around skewed data and reward hacking, tricks that earn rewards without true answers, which show up only in long tool-using runs. If Reflection kept quality high to 100 million rollouts, that testing system is the lasting asset. It would carry over to later models better than any single model file, and it would give company teams an open option built for careful, multi-step work rather than one-shot replies.


