AI Technology Observations | 6 October 2026, 07:00

This observation covers the period from 7:53 am on 5 October to 7:53 am on 6 October 2026 in Sydney, Australia. It records one formally announced AI technology change that can be cross-checked against a first-party source and independent reporting.

Reflection details Beam while the weights remain pending

Reflection AI formally introduced Beam on 5 October and made an early version available to selected users. The company describes it as a text-only sparse mixture-of-experts model for coding, reasoning and agentic workloads, with 501 billion total parameters, 23 billion active parameters per inference step and an effective context length of one million tokens. Compared with the earlier notice that a model was approaching release, the new material provides a combined account of the architecture, training scale, evaluations, access arrangements and planned open release. The current product state is still selected-user early access, not a complete public release of the weights.

According to the developer, Beam was pretrained on 23.8 trillion tokens and then underwent high-compute reinforcement learning involving more than 100 million rollouts. Reflection says pretraining took less than four weeks on 6,144 NVIDIA GB300 NVL72 GPUs, while reinforcement learning ran for four weeks on 10,500 GB300 GPUs and used about 1.3 billion sandbox environments. These training-scale and efficiency figures are company-reported and have not been independently audited or reproduced.

Reflection also published results for evaluations including SWE-bench Verified, Terminal-Bench 2.1 and GPQA Diamond, and said Beam used less inference compute than several comparison models on some tasks. TechCrunch independently confirmed the reported parameter count, training-token total, context length and release arrangements, while noting that the benchmark claims had not been independently verified. The public material does not present Beam as a multimodal model; its currently stated scope is text, coding, reasoning and agentic workloads.

At the end of this observation period, Beam was still undergoing final red-teaming and evaluation. Reflection plans to release the weights under the Apache 2.0 licence later in October, together with a technical report, model card, safety-evaluation results, documentation, and its training and inference stack. Those materials are not all public yet. “Open-weight” therefore describes the announced release plan, not a delivery completed in this observation period.

Sources

Reflection AI, Introducing Beam, 5 October 2026
https://reflection.ai/blog/introducing-beam

TechCrunch, Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost, 5 October 2026
https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/