Reflection AI Beam is announced as 501B open weights, but Hugging Face lists no files

Reflection AI announced Beam on October 5, 2026, calling it the company’s first open-weight model: a sparse mixture-of-experts design with 501 billion parameters in total and 23 billion active for each token. Reflection AI Beam was promised under the Apache 2.0 licence, with the weights, a technical report and a model card due “later this month”, the announcement says. Ten days later, none of that can be downloaded. On October 15 the verified Hugging Face organisation whose card calls it Reflection’s official account listed no models, and access to Beam runs through a waitlist for an early version that Reflection says is going to “a select group of users”. Every figure below is Reflection’s, taken from its announcement and its developer documentation, plus Hugging Face and GitHub data we saved on October 15.

Takeaway points

• Reflection announced Beam on October 5, 2026, as a mixture-of-experts model with 501 billion parameters in total and 23 billion active, and said it would release the weights under Apache 2.0 this month.
• On October 15 Reflection’s verified Hugging Face organisation listed 0 models, so the weights, the model card and the technical report could not be downloaded or read, and access was through a waitlist.
• We have not run Beam, and with no weight files published there is no download size to set against the 48 GB of memory on our test machine.

What is Reflection AI Beam and when was it announced?

Beam is a text-only language model that Reflection announced in a blog post dated October 5, 2026, titled “Introducing Beam: Reflection’s 501B open-weight model”. The post describes it as a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic work. In a mixture-of-experts model each token passes through a few of many sub-networks, so only part of the model does the work for any one token. Reflection’s developer documentation lists the model under the identifier Beam-1.0-501B-A23B, with a creation date of October 5, 2026.

The announcement says Beam “is undergoing final red-teaming and evaluations” and that early access is by sign-up. Its closing section says the early version is going to a select group of users through a waitlist. The developer documentation says the same of the platform as a whole: it is in beta, and access is opening gradually through a waitlist.

The documentation gives these figures for the hosted model.

The documentation adds that the default effort shown is an example and can differ between organisations. It also says Beam supports reasoning, tool calling and structured outputs.

Can you download Beam from Hugging Face yet?

No. On October 15, 2026, at 13:58 UTC, the Hugging Face models listing for the organisation named reflection returned an empty list, and the organisation’s overview counted 0 models and 0 datasets. That organisation is verified, carries the full name Reflection, and its card says it is the official Hugging Face organisation for Reflection, linking to reflection.ai and to the company’s developer documentation. Reflection’s own website and the Beam announcement do not link to any Hugging Face page.

We also checked the other account names the company might use. The listings for reflection-ai and ReflectionAI returned nothing, and a Hugging Face search for Beam-1.0 returned no repository. The GitHub organisation linked from the Hugging Face card, reflectionai, held 12 repositories on October 15, and none had Beam in its name.

The announcement makes two promises about timing, both in its own words. Near the top it says “We will release the weights, technical report, model card, and developer artifacts later this month.” In its closing section it says “This month, we will release the weights under an Apache 2.0 license”. Read against the post’s date, both point to October 2026. Until the files appear there is no licence file, no config file and no parameter count from the safetensors files to check the 501 billion figure against.

That check has mattered before. When DeepSeek published its weights, DeepSeek V4.1 Flash weights are out: 552B in the announcement, 763.2B in the files found a gap between the two numbers. For Beam no such comparison is possible yet.

How does Beam score against GLM 5.2 and Kimi K3, by Reflection’s figures?

By Reflection’s own table, Beam is within one point of GLM 5.2 on four of the six rows below, and behind Kimi K3 on all five of these rows where Kimi K3 has a score. Reflection marks scores that were not reported as NR. The post says the tables were updated on October 8, 2026, and that all other results are unchanged. We have not reproduced any score in this section.

The post says plainly that Kimi K3 remains ahead on raw capability and that Beam’s case is efficiency at inference time. It claims Beam reaches scores comparable to GLM 5.2 on advanced reasoning tests while using three to four times less inference compute. That comparison rests on an estimate, not a measurement. Reflection says it worked out compute as two times the active parameter count times the mean number of generated tokens, and that the estimate leaves out prompt prefill, context-dependent attention and serving overhead. It says the scores for other models come from Artificial Analysis and DataCurve.

What do Reflection’s own documents disagree on about Beam?

They disagree on context length. The announcement says midtraining “extends Beam’s effective context length to 1M tokens.” The developer documentation lists a context window of 256K for Beam-1.0-501B-A23B, and its example response gives a context length of 262,144 tokens, with a note that the window may change during the beta.

The two can both be true, since one describes training and the other the hosted service. Neither document says so, and there is no model card yet to settle which figure a downloaded copy would support. The reinforcement learning stage, the announcement says, used rollouts with a maximum context length of 256K tokens.

The announcement also gives the size of its training task pool in two ways. In one passage Reflection says it sourced “one million” coding, agentic and STEM environments; in another it says it built “a pool of nearly one million environments”. The gap is small, and we report it only as the wording stands.

How was Beam trained, according to Reflection?

Reflection says it pretrained Beam on 23.8 trillion tokens from the web, public sources and licensed datasets, in under four weeks on a cluster of 6,144 Nvidia GB300 NVL72 GPUs. The post says the run needed nine semi-automatic rewinds, which it puts down to gradient norm spikes or suspected silent data corruption, and that goodput reached 92.3% towards the end. A chart caption in the post says Beam has 52 layers.

For reinforcement learning, the post says it used 10.5K Nvidia GB300 GPUs for four weeks and generated more than 100 million rollouts, with training and grading using about 1.3 billion sandboxes. It says capabilities kept improving as reinforcement learning compute grew, “with no sign of a plateau”. That is Reflection’s claim, and the evidence for it is a set of charts in the post rather than a published report.

On safety, the post says Reflection trained a separate safety and alignment model from the same pretrained checkpoint and merged the two through distillation. It says the results of its safety evaluations will be published in the technical report.

What has Reflection not published about Beam yet?

Reflection has not published the weights, the model card or the technical report that its October 5 announcement promised for this month. As of October 15 there was nothing to download from its Hugging Face organisation, so the 501 billion parameter count, the Apache 2.0 licence and the context length cannot be checked against files. The safety results are also held back for the technical report, and the post does not name the distribution partners it says Beam will launch with.

Size is the open question for anyone hoping to run it locally. When weights are published, the size of the files is the first figure to set against a desktop machine’s memory, as in Aleph Alpha Kolibri is a 78.84 GB open-weight model, and its own documents disagree on German data. Beam has no repository, so no such figure exists. Reflection has not said whether it will publish quantised builds or which file formats it will use.

We have not run Beam. Our test machine is an Apple M4 Pro with 48 GB of unified memory, and with no published weights there is no file size to compare with it and nothing to load. This piece gives no speed, memory or quality figure of our own. For another large open-weight release we have reported on, see IQuest releases IQuest-Q1, a 320 billion parameter open weights model, on September 28.