Beam, the West's answer to China's open models, arrives without its weights

The announcement is unusually detailed. Reflection AI says it pre-trained Beam on 23.8 trillion tokens drawn from the web and from licensed proprietary datasets, using 6,144 NVIDIA GB300 NVL72 GPUs for under four weeks at 92.3% goodput; then came a reinforcement learning phase with more than 100 million rollouts on 10,500 GB300s for another four weeks, roughly 1.3 billion sandboxes and around a million coding, agentic and STEM environments. Native context is 256K tokens, extended to one million through mid-training — a distinction the blog makes and that TechCrunch, which reports only the one-million figure, doesn't pass on. Numbers this specific have an effect: they read like verification, when they're a claim. TechCrunch notes that nobody has independently checked the claimed performance.

The central claim is efficiency: reasoning performance comparable to Z.ai's GLM-5.2 with '3-4 times less' inference compute. It's worth reading the methodology note Reflection itself publishes: compute is estimated as FLOPs ≈ 2 × active parameters × average tokens generated per attempt, excluding prefill, context-dependent attention operations and serving overhead. That's arithmetic on active parameters, not a measurement of cost or latency on a real server — and the excluded items weigh on the real cost for whoever runs the model, which is why the estimate doesn't tell you how much you actually save.

The benchmark table is more interesting than a launch announcement would require, because it also shows the losses. Beam sits below several Chinese open models: SWE Bench Pro v2-Hard 77.2 versus 84.3 for GLM 5.3 and 88.2 for Kimi K3; Terminal Bench v2.1 80.1 versus 81.0 for GLM 5.2 and 90.6 for DeepSeek V4.1; HLE without tools 36.2 versus 46.9 for Kimi K3; GPQA Diamond 90.5 versus 93.5 for Kimi K3. Where it wins clearly is against Western models: against Inkling on nearly every item compared, from SWE Bench Pro v1 at 65.5 versus 54.3 to Terminal Bench v2.1 at 80.1 versus 63.8, and against Nemotron 3 Ultra on most tests. Several cells read 'NR', scores not reported for competitors, so the comparison is partial by construction.

Reflection pitches Beam as a 'workhorse' model for everyday use by enterprises, the public sector and developers, distributed through hyperscalers, neoclouds and integrations into open-source libraries. According to TechCrunch, the company, founded in 2024, raised about $4.7 billion at a $25 billion pre-money valuation in its latest round, counts Nvidia, Sequoia Capital and Lightspeed among its investors, and has compute deals worth more than $7 billion in total with SpaceX and Nebius for GB300 chips through 2029.

What I'm watching is the gap between what's public today and what's promised. Today there's an unsigned blog post, a table compiled by the company and a waitlist on platform.reflection.ai; the weights, the Apache 2.0 license, the documentation and the stack for running, evaluating and fine-tuning are all commitments dated 'later this month'. Meanwhile, the Chinese open models Beam measures itself against are already available, and anyone can redo the math on them. It's a difference no benchmark records: publishing your weights means handing others the tool to prove you wrong. Until then, the sentence to use isn't 'Beam is competitive' but 'Reflection says Beam is competitive' — and the distance between the two is measured in October days.

— Olya