What Reflection's Beam post schedules

What happened

On October 5, 2026, Reflection published a post introducing Beam and calling it the company's first open-weight model. The post describes a sparse mixture-of-experts model with 501 billion parameters in total and 23 billion active, built for coding, reasoning, and agentic work. The same post says the weights, the technical report, the model card, and the developer artifacts are scheduled for later this month, under an Apache 2.0 license. Until that release, the company is offering early access through a waitlist, and it says final red-teaming and evaluations are still underway. A reader who treats the words open-weight as a file that can be fetched today is ahead of the post. The page filing that post is Futureweb.

What the post already counts

Pretraining, the post says, used 23.8 trillion tokens and finished in under four weeks on a cluster of 6,144 NVIDIA GB300 NVL72 GPUs. A separate reinforcement-learning run used 10,500 GPUs of that class for four weeks and generated more than 100 million rollouts, with a maximum context of 256,000 tokens during the run. The post says that run used about 1.3 billion sandboxes and a pool of nearly one million environments. Midtraining, it says, extends the model's effective context to one million tokens. Those figures are the company's. The weight file that would let a second lab repeat them is the release that has not happened.

The license is still a date

Apache 2.0 is a named license, and the post puts it on the weights later this month, together with documentation and a stack for running, evaluating, and fine-tuning. The post also says Beam will ship with distribution partners and with integration into open-source libraries. Until a reader can fetch the weights, the license is a sentence in the announcement, not a file a second lab can hash. What a license obligates, once a file exists, is the standing note on what a license obligates.

Why it matters on this desk

This desk files what a model release actually includes. Beam's post includes a parameter split, a training description, a waitlist, and a date for the weights. It does not include the weights. The company also says that on some reasoning benchmarks Beam uses three to four times less inference compute than GLM 5.2, and that the comparison is an estimate. The post says the estimate excludes prompt prefill, context-dependent attention, and serving overhead, and that it is not a measured serving bill. A second lab cannot rerun that estimate from a public checkpoint on October 6. The two context figures in the post, 256,000 tokens in the reinforcement-learning run and one million from midtraining, are the kind of bound filed under what a context window actually bounds.

Where the accounts add a ranking

TechCrunch, The Decoder, and MarkTechPost all covered the October 5 post. The Decoder's headline calls Beam the most capable open-weight model built outside China. Reflection's post says Beam advances the Western open-weight frontier, is competitive with larger open models such as GLM 5.2 on coding and agentic tasks, and is approaching Qwen 3.8-Max on those tasks, while models such as Kimi K3 remain ahead on raw capability. The advantage the company claims is efficiency at inference time. Most capable is the outlet's ranking. The post's ranking is narrower, and it is still the company's own table. This page will not adopt either line as a result a reader can reproduce before the weights ship.

What a reader can check

The check is the October 5 post. Are the weights linked for download? The post says later this month. Is the license named? Apache 2.0, on that future release. Are 501 billion total and 23 billion active in the opening? Yes. Is a second lab's evaluation linked? No. The company says it will publish safety-evaluation results in the technical report and will open-source safety evaluations it used internally. That report is part of the same later-this-month package. What stays on the device once a weight file can actually be run is the evergreen on what on-device inference keeps.

What to watch next

The page changes when Reflection publishes the weights, the technical report, and the model card, or when it changes the license named on October 5. A benchmark table that appears only on an outlet, with no weights behind it, does not replace that release. Early access behind the waitlist is the access the post offers now. A later note that the weights slipped past this month, or that the license is no longer Apache 2.0, would replace the sentence in the opening. Until then the post is the record, and the outlets are accounts of the post.

What would change the record

A public weight URL, a model card a reader can open, or a withdrawal of the Apache 2.0 sentence would change the sentences above. An early-access demo that stays behind the waitlist does not. The post says the preview shows what Beam can do today for a select group of users. A claim that the model is fully released on October 6 is not what the post says. The training-cluster sizes and the token count stay the company's account until the technical report gives another lab the artifacts to test them.

What stays on this desk

What stays specific is the split between the announcement and the file. Beam is described. The weights are dated for later this month. A radio license, a certificate authority, and a handset rumor are not this post. The desk for the announcement is Artificial Intelligence.