Reflection sells Beam to sovereign-AI buyers on a 3-to-4x inference compute claim

The efficiency figure appears in two versions. RuntimeWire, relaying Semafor, reports Laskin saying Beam needs three to four times less computing power than comparable open models to reason through a problem ^[3]. A dev.to account of the launch attaches the same range to inference compute per generated token ^[5]. Those are different measurements. Per-token compute in a mixture-of-experts model follows the active parameter count, and Beam uses about one parameter in 22 on each token ^[22]. A per-problem figure also depends on how many tokens the model spends reasoning before it answers.

Hardware is sized by a different number. The model activates only some experts on each token, but all 501 billion parameters stay loaded in memory ^[6]. If the weights ship at 16-bit precision, they come to about 1 TB before any context cache ^[24]. RuntimeWire's review found that Reflection's compute comparison leaves out several costs of serving a model ^[12]. Neither report states the hardware, batch size or precision behind the three-to-four-times figure.

In my view the ratio transfers to a buyer only if that buyer's serving setup resembles the one Reflection measured, with similar batch sizes and output lengths. Until the technical report defines the comparison ^[1], I would size hardware for the full parameter count and treat the throughput gain as a company claim.

The training disclosure is more concrete, and the engineering behind it is serious. The four-week reinforcement-learning run works out to about 7.06 million GPU-hours ^[25]. To grade the rollouts, Reflection ran roughly 1.3 billion sandboxes fed by one million coding, agent-use and science environments ^[18]. The company calls it one of the largest RL runs reported by an open lab ^[20]. Those rollouts used contexts of up to 256,000 tokens ^[18], about a quarter of the advertised window ^[23].

On capability, Reflection says Beam performs comparably to Z.ai's GLM-5.2 on advanced reasoning benchmarks ^[4]. By the dev.to account, the company also places it close to Qwen 3.8-Max, concedes that Kimi K3 is ahead on raw capability and argues Beam's edge is inference efficiency ^[10]. For those scores to predict production results, a buyer's workload has to resemble the benchmark suite.

Laskin told Semafor that customers looking to build sovereign AI systems "don't really have very good options today" ^[7]. Reflection's wider plan is an "AI factory" for customers that train and deploy their own systems in private clouds, on premises or in restricted environments ^[13]. Most of the open models leading benchmark tables come from Chinese labs, dev.to notes, and Reflection frames Beam as an open option trained outside that group ^[19]. Its own comparison set also includes Nemotron 3 Ultra and Inkling, Thinking Machines' mixture-of-experts model ^[11]. Neither is among the Chinese-lab families dev.to names ^[19]. Reflection's June financing valued the company at $25 billion pre-money, according to Semafor ^[15].