Reflection publishes Beam weights, a 501-billion-parameter model under Apache 2.0
Hugging Face, Friday: Beam-1.0-501B-A23B is public in BF16, FP8 and NVFP4. About 23 billion parameters active, 256 experts with 8 routed, a 1 million-token window, text only, Apache 2.0. A Rust harness ships with it, and Transformers, vLLM and SGLang support it on day one. Reflection announced the model on 5 October. S5 Labs, updating on 16 October, says the technical report is still marked as coming.
Feature3 minIn-depth
James Okonkwo3 min read
- Reflection
- Beam
- Open weights
- Hugging Face
- Agents

Photo: ReflectionSource page
Hugging Face on Friday published the weights for Beam, Reflection’s first open model. Ben Burtenshaw and Pedro Cuenca, in a post dated 16 October, described a text-only mixture-of-experts system with 501 billion parameters in total, about 23 billion active for each token, a context window of 1 million tokens, and an Apache 2.0 license. The checkpoint is on the Hub as reflection/Beam-1.0-501B-A23B, in BF16, FP8 and NVFP4. The model card lists 501.3 billion parameters in total and 23.4 billion active, 52 layers of which the first three are dense and the other 49 are mixture-of-experts, 256 experts in each of those layers, and eight experts routed to a token. The knowledge cutoff on the card is 30 June 2026. The post says the model is built for coding and agentic work, that it ships with a dedicated open-source harness written in Rust, and that Transformers, vLLM and SGLang support it from the day of release.
The weights were promised, not shipped, eleven days earlier. Reflection’s post on 5 October introduced Beam as a sparse mixture-of-experts model with 501 billion parameters and 23 billion active, pretrained on 23.8 trillion tokens in under four weeks on 6,144 NVIDIA GB300 NVL72 GPUs. The reinforcement-learning run, it said, produced more than 100 million rollouts on 10,500 GB300 GPUs over four weeks. That post said the weights, a technical report, a model card and developer artifacts would come later in the month. Joshua Wendt at S5 Labs, updating a 5 October article on 16 October, wrote that the weights are now public and ungated, that the repository’s license file is the standard Apache 2.0 text, and that the technical report was still marked as coming. He also wrote that he could not establish the day the repository became public. The model card is up. The report, as of that update, is not.
Scores, and whose they are
The numbers on the card are Reflection’s, run at temperature 0.7 and top_p 0.99. On the BF16 checkpoint the card lists SWE-bench Verified at 81.6 percent, Terminal-Bench 2.1 at 80.1 percent and GPQA Diamond at 90.6 percent. S5 Labs noted two cells that do not match the 5 October post: SWE-bench Pro v2-Hard is 76.5 percent on the card and 77.2 in the post, and Humanity’s Last Exam without tools is 36.0 against 36.2. The October post compared Beam with larger open models and said it was competitive with GLM 5.2 and behind Kimi K3 on raw capability, with the advantage in inference compute. Those comparisons are Reflection’s table. No independent score is in the posts that went up with the weights.
Serving it is a memory question before it is a speed claim. Hugging Face said the BF16 checkpoint is about 1 terabyte in safetensors, and that the NVFP4 build, which keeps some weights in BF16, brings that to about 280 gigabytes for NVIDIA Blackwell GPUs. A B200 has 180 gigabytes of memory, so two of them can hold the NVFP4 weights with little room left for a cache. The post says to plan on four B200s. FP8 is the checkpoint it points to for GPUs older than Blackwell. The Rust harness is not given a product name in the Hugging Face post. It is described as the harness that gets the best performance out of the model.


