Reflection has made the weights of its Beam model publicly available, ten days after previewing it. Beam is now on Hugging Face under the Apache 2.0 licence, which lets anyone download, modify and use it, including commercially.
The release is described in a Hugging Face blog post published on 16 October and written with Reflection. It calls Beam “a mixture-of-experts (MoEs) model with 501B parameters (23B active per token), supports text as input, has a context window of 1M tokens, and is released under the Apache 2.0 license.” The weights are in the reflection/Beam-1.0-501B-A23B repository, and there are separate FP8 and NVFP4 versions. Reflection’s blog also says “Beam is now available”.
What Reflection has released
Beam is a mixture-of-experts model. Instead of using all of its parameters for every word it processes, it sends each token to a small group of specialist sub-networks, called experts. According to Hugging Face, the model has “52 layers, 256 experts per MoE layer, 8 routed per token.” As a result, only 23 billion of its 501 billion parameters are used for each token.
That matters for anyone running it. Hugging Face explains that all the experts still have to be held in memory, so deploying Beam needs the memory of a 501-billion-parameter model, but it runs at roughly the speed of a 23-billion-parameter one.
Beam takes text in and gives text out. Its context window of one million tokens means it can take in very long documents or large codebases in a single request.
What it takes to run
Beam is not a model for a laptop. Hugging Face says the full-precision BF16 checkpoint is “about 1 TB”. The NVFP4 version, which stores most of the parameters at 4-bit precision for Nvidia’s Blackwell GPUs, “shrinks the 501B parameters in BF16 from 1 TB to ~280 GB.”
Even that smaller version needs data-centre hardware. Hugging Face says two Nvidia B200 GPUs can just about hold the weights, but with little room left for long contexts. Its advice is: “Plan for four B200s (720 GB).”
Beam works from launch with the widely used transformers, SGLang and vLLM libraries. It is also available through Hugging Face Inference Providers, “powered by Baseten, Fireworks, and Together”, so developers can try it without their own hardware. Hugging Face says Beam ships with “its own dedicated open source Rust based harness”, and that it has published recipes, worked out with Reflection, for running the model and for fine-tuning and reinforcement learning. To encourage people to build with it, Reflection will give “$10,000 Hub credits to the best projects that use Beam”.
The benchmarks are Reflection’s own
The Hugging Face post includes benchmark scores “as provided by the Reflection team”. They have not been independently checked. For the BF16 model, Reflection reports:
- SWE-bench Verified (fixing real software bugs): 81.6%
- SWE-Bench Pro v1: 66.1%
- Terminal-Bench v2.1 (working at a command line): 80.1%
- GPQA Diamond (graduate-level science questions): 90.6%
- AIME 2026 (competition maths): 97.7%
- Humanity’s Last Exam, without tools: 36.0%
The table also gives scores for the FP8 and NVFP4 versions, which are close to the BF16 figures. Until independent groups test Beam, these results should be read as Reflection’s claims.
Why Beam’s training run and licence matter for self-hosting
Reflection’s launch post gives more detail on how Beam was built. The company says it pretrained the model on 23.8 trillion tokens from the web and proprietary licensed datasets, then ran a reinforcement learning stage that generated over 100 million rollouts on 10,500 Nvidia GB300 GPUs over four weeks. It pitches Beam on efficiency at inference time rather than raw capability, and says larger open models such as Kimi K3 remain ahead on capability.
That pitch fits the release. A model with 23 billion active parameters answers more cheaply than a dense model of similar total size, and the Apache 2.0 licence lets a company run it on its own servers, fine-tune it on private data and ship products without paying a licence fee. For banks, hospitals or public bodies that cannot send sensitive data to an outside API, a capable open model they control is a real option. The catch is the hardware bill: four B200s is a data-centre purchase, not a test bench. Companies weighing that choice also need a plan for securing and monitoring the model, which Microsoft and Nvidia set out in their risk guide for open-weight models in enterprise. Smaller teams are likely to use the hosted providers first.
From a waitlisted preview to public Beam weights
TSN covered Beam’s preview on 6 October. At that point the weights were available only through a waitlist, and an Apache 2.0 release was promised for “later in October”. The public release is what has now happened.
Reflection has also been the subject of deal speculation. On 10 October TSN reported that, according to the Financial Times, Nvidia was in talks to acquire or deepen its investment in Reflection. Neither company has announced a deal.
Update, 1 November 2026: VentureBeat reported on 16 October that Reflection’s chief technology officer, Ioannis Antonoglou, called the acquisition reports “baseless rumors”, describing Nvidia as a close partner and investor and saying Reflection intends to stay independent. No independent benchmark results for Beam had been published in the sources we checked.
Independent Beam benchmarks and hosting prices are still missing
- How Beam performs in independent tests, compared with Reflection’s own figures.
- What the hosted inference providers will charge for Beam.
- Whether there is any agreement between Nvidia and Reflection. The FT report has not been confirmed by either company.
Fine-tunes, leaderboards and the Nvidia question
- Independent benchmark results and leaderboard placings for Beam.
- Fine-tuned versions and smaller derivative models appearing on Hugging Face.
- Further hosting partners, and any announcement from Reflection or Nvidia about the reported talks.
Beam’s Apache 2.0 licence, hardware needs and unchecked scores
- Beam is now openly available under a permissive licence, so companies can host and adapt it themselves.
- Because the model is so large, running it needs several high-end GPUs or a hosted provider.
- The benchmark claims have yet to be tested by anyone outside Reflection.
Sources
- Hugging Face, “Welcome Beam: a 501B open model you can own, customize, and scale”, 16 October 2026
- Hugging Face, reflection/Beam-1.0-501B-A23B model card
- Reflection blog
- Reflection, “Introducing Beam: Reflection’s 501B open-weight model”, 5 October 2026
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- Reflection’s Beam: an open-weight workhorse pitched on efficiency
- Nvidia Reportedly in Talks to Acquire or Deepen Its Investment in Reflection AI
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X





