Reflection Beam, the first open-weight model from Reflection AI, had its weights released on Hugging Face under the Apache 2.0 license on October 16, 2026. The model is a Mixture-of-Experts (MoE) with 501 billion total parameters and about 23 billion active per token, designed for coding, reasoning, and long-horizon agentic workloads.
What was launched
According to the company's official announcement, the weights are available in BF16, FP8, and NVFP4 formats, along with a drafter model for speculative decoding. Context reaches 1 million tokens. Reflection describes Beam as the kernel of an "open operating system for agents," with integrations across more than 40 partners, including AWS, Google Cloud, NVIDIA, and tools like vLLM and SGLang.
The model is text-only and was trained predominantly in English, with a knowledge cutoff of June 30, 2026. The company emphasizes efficiency: reports indicate it achieves results comparable to models like GLM-5.2 using fewer inference tokens.
Reported benchmarks
In Reflection's own numbers (temperature 0.7), Beam (BF16) scores 81.6% on SWE-bench Verified, 80.1% on Terminal-Bench 2.1, 66.1% on SWE-Bench Pro v1, and 90.6% on GPQA Diamond. Quantized versions maintain close performance.
These results place it competitively with Chinese open-weight models on some agentic metrics, though still behind leaders like DeepSeek V4.1 Flash on other tests. Reflection positions Beam as advancing the Western open-weight frontier, prioritizing cost per completed task over maximum raw score.
Context and availability
The preview was announced on October 5, with weights promised for the same month. The release coincides with other recent openings of large models for agents, such as Ant Group's Ling-3.1-flash (MIT) and StepFun's Step 5 Preview. Developers can download the weights, use Reflection's API (free beta to start), or access via partners.
The Apache 2.0 license allows commercial use and modifications with few restrictions, unlike some custom licenses of other open-weight models.
Sources
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult original sources for full context.
By GeekikiBot