On October 17, 2026, Beijing time, Reflection officially released the instruction fine-tuning weights, model cards, and security technical reports for Beam, making them available for download under the Apache 2.0 license. The official model card lists Beam as a sparse mixture-of-experts model capable of accepting text inputs and producing text outputs, with a total parameter count of 501.3 billion and 23.4 billion parameters activated per token. The model's official news page linked the weight entry to Hugging Face on the same day, officially launching it rather than merely previewing it. Reflection positions Beam for software engineering, programming, reasoning, and long-horizon agent tasks; however, existing standalone materials have not validated the company’s overall assessment regarding efficiency and capability.
Weight Licensing and Model Architecture
Beam weights are simultaneously provided in three numerical formats: BF16, FP8, and NVFP4. The release also includes a draft model for speculative decoding. The Apache 2.0 license applies to the weights released by Reflection, but this does not imply that cloud services, managed interfaces, or commercial products from more than forty partners are open-source. Developers can download, self-host, evaluate, and customize the weights, or access them through Reflection’s interfaces and those of its partners. Official materials only list the text modality and do not describe Beam as a multimodal model capable of processing images, audio, or video. Different precision formats do not indicate three distinct models with separate knowledge bases or functional boundaries.
The architecture listed in the model card includes 52 layers: 3 dense layers and 49 mixture-of-experts layers, with a total of 256 experts. The attention structure employs gated grouped-query attention and sliding window attention. The officially announced maximum context window is 1 million tokens, with a vocabulary size of 152,000 tokens and a knowledge cutoff date of June 30, 2026. These are specifications for the model and do not mean that every test was conducted using a 1-million-token context window, nor should the total parameter count of 501.3 billion be misinterpreted as all parameters being simultaneously involved in each generation.
Launch Channels and Task Attribution
Reflection stated that at the time of launch, there were already over 40 partners and integrations in place, with weighted entry points connecting to Hugging Face. Their own APIs are in public beta and can be used for free to begin with, though the official pricing has not yet been announced. The official list covers cloud platforms including AWS, Google Cloud, and Microsoft Azure, as well as several managed inference providers, model serving tools, and agent development environments. Dell and HPE were listed as deployment paths behind enterprise firewalls. Reflection also listed Transformers, vLLM, SGLang, and Ollama as first-day supported tools; the functionality, hardware requirements, and pricing of these channels differ.
Reflection describes Beam's primary use cases as software engineering, agent programming, long-range tasks, general inference, and subsequent fine-tuning. According to model cards, under settings with temperature 0.7 and top-p 0.99, SWEBench Verified scores were 81.6% for BF16, 80.2% for FP8, and 80.7% for NVFP4. These figures come from Reflection's release materials, indicating results for the three weight formats in this particular test; they do not cover all tasks, service frameworks, or hardware, nor are they independent cross-model rankings.
Performance Claims and Safety Assessment
Efficiency is the core claim of Reflection's latest release, with the company stating that Beam achieves comparable problem-solving quality with fewer reasoning tokens. Google Cloud announced same-day deployment support and claimed that in customer benchmarks on A4X Max instances, llm-d, or NVIDIA Dynamo service stacks, throughput could increase up to 5 times and latency could decrease by up to 8 times. Google Cloud explicitly limited these figures to customer benchmarks, meaning they cannot be stated as guaranteed improvements for the Beam model itself across any cloud, on-premise device, or workload.
Reflection's security report states that Beam's training combined pre-training data filtering, supervised fine-tuning, reinforcement learning, and multi-teacher policy distillation. The company rated the marginal impacts of four categories of systemic risks - biological weapons, chemical weapons, loss of control, and cyber attacks - as Limited according to the Open Safety Framework. Its Iris dialogue evaluation records show that the BF16 version had an 88.9% refusal or redirection rate for harmful requests and an 81.5% compliance rate for harmless requests; these remain company assessments under Reflection's defined methodology and cannot be converted into absolute safety rates for real-world deployments.
Caliber Discrepancies and Information Gaps
The sources do not conflict over whether the weights have been released, but they use different levels of numerical precision: the release page rounds the scale to about 501 billion total parameters and 23 billion activated parameters, while the model card lists 501.3 billion and 23.4 billion; this article uses the latter values. VentureBeat's public metadata confirms the headline, author Matt Marshall, and a publication time of 6:00 AM Pacific on October 16, or 1:00 PM UTC. That metadata was not used to add technical details, and it does not independently reproduce Reflection's or Google Cloud's performance and safety figures.
As of publication, Reflection had not disclosed official post-beta pricing, paid tiers, actual user numbers, long-term operating costs, or unified prices across partner channels; free access during the public beta cannot be interpreted as a promise that the service will remain free. The available materials also do not provide independent cross-platform evaluations covering different hardware, caching strategies, concurrency levels, prompt lengths, and service prices; long-term safety performance in real environments similarly lacks public operating data. What can be confirmed is that the model weights, three numerical formats, model card, safety report, and multiple deployment paths were all made public on the same day.