Reflection · siet 5. Oktober 2026 (Ankündigung); Modellgewichte für später im Oktober 2026 angekündigt · 17× · tolest 07. Okt. 2026
Beam is Reflection AI's first open-weight language model, unveiled on October 5, 2026. It is a text-only sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active per token, built for coding, reasoning, and agentic workloads. The model was pretrained on 23.8 trillion tokens and, in its current API beta, supports a context of up to 256,000 tokens (Reflection states midtraining extends the effective context to up to 1 million tokens). Weights are planned for release under an Apache 2.0 license later in October 2026; until then, access is limited to a waitlist-gated beta API with no published pricing.
Features
| Key Benchmark (%) | SWE-bench Verified: 80.9% (self-reported by vendor, not independently verified) |
| Context Window (Tokens) | 256,000 tokens in beta API; per Reflection, effectively extended to up to 1,000,000 tokens via midtraining |
| License | Apache 2.0 (planned, upon weights release later in October 2026) |
| Multimodality | Text-only, no native image, audio, or file inputs |
| Platform | Beta API (api.reflection.ai, OpenAI-compatible), waitlist access; weights planned later for self-hosting |
| Price per 1M Tokens | Not published (as of October 2026) |
| Release Date | October 5, 2026 (announcement); weights to follow later in October 2026 |