On October 5, 2026, Reflection AI officially unveiled Beam, its first frontier open-weight model. The company, founded about two years ago, described the launch in a lengthy blog post that confirms what Axios had reported over the weekend: the startup was close to showing a model with which it hopes to build a Western answer to Chinese open models such as DeepSeek, Qwen and those from Z.ai.
Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, of which only 23 billion are active for each token. That design lets the model store a great deal of capacity while keeping inference costs comparable to much smaller systems. It is a text-only model focused on coding, reasoning and agentic tasks.
According to Reflection, pretraining used 23.8 trillion tokens drawn from the web, public sources and proprietary licensed datasets. The company says it eliminated about 95% of raw internet tokens through parsing, deduplication and filtering, and that the full pretraining run was completed in under four weeks on a cluster of 6,144 NVIDIA GB300 NVL72 GPUs.
The most striking part of the announcement is the bet on large-scale reinforcement learning. Reflection says it deployed 10,500 NVIDIA GB300 GPUs for four weeks to generate more than 100 million practice rollouts, with contexts of up to 256,000 tokens, drawing on nearly one million coding, agentic and science environments. The company argues that capabilities kept improving as reinforcement compute increased, with no sign of a plateau.
On results, the figures published by the company put Beam at 80.9 on SWE-bench Verified, 80.1 on Terminal Bench v2.1, 90.5 on GPQA Diamond and 44.4 on DeepSWE v1.1. Reflection presents it as competitive with larger open models such as GLM 5.2 and close to Qwen 3.8-Max on coding and agentic tasks, while acknowledging that models like Kimi K3 remain ahead on raw capability.
Reflection's central argument is efficiency. The company claims that on advanced reasoning tests Beam achieves scores comparable to GLM 5.2 while using three to four times less inference compute. To get there, it trained the model with a controllable length penalty and offers a reasoning-effort parameter, so users can choose shorter, cheaper answers or longer, more accurate ones.
It is worth stressing, as TechCrunch did, that these numbers have not been independently verified. They are results reported by Reflection on its blog, with comparisons partly based on third-party data from Artificial Analysis and DataCurve, and with several cells marked as “not reported.” External validation will come once the weights are available and the community can reproduce the tests.
That moment is close. Reflection says Beam is in the final stage of red-teaming and evaluations and that this month it will release the weights under an Apache 2.0 license, along with a technical report, a model card and tooling to run, evaluate and fine-tune it. In the meantime, a preview version is available to a select group of users through a waitlist.
On safety, the company explains that it trained a second model dedicated to behavioral principles and then merged the capabilities of both through multi-teacher distillation. Its rules are organized in three tiers: rules the model must not break, qualities it should consistently satisfy, such as acknowledging uncertainty, and a direct, efficient default style. Reflection promises to publish its safety evaluations in the technical report.
The launch has an obvious geopolitical dimension. Over the past year, the most capable open models have mostly come from Chinese labs. Reflection positions itself against them and also against Western players such as Mistral, Meta and Inkling, the open model from Mira Murati's Thinking Machines Lab released in July. According to Reflection's own tables, Beam beats Inkling on several coding tests, although Inkling is multimodal and Beam is not.
For companies, public agencies and developers, an open model with a permissive license means being able to run it on their own infrastructure, fine-tune it with internal data and avoid dependence on a single provider. Reflection calls Beam a “workhorse” model for enterprise workloads and says it will launch with distribution partners on major clouds and integration into open-source libraries.
The challenge, however, will be practical. A 501-billion-parameter model requires substantial hardware even if it activates only a fraction per token, and the claimed efficiency will need to be confirmed in real deployments. Competition is also intense: on October 6, Mistral announced a preview of its own one-trillion-parameter open model.
In short, Beam makes Reflection a serious contender in the race among Western open models, with a clear emphasis on reinforcement learning and efficiency. If the weights ship on schedule and independent testing backs up its numbers, the model could become a relevant option for anyone seeking coding and agent capabilities without relying on closed models.
