Unknown · depuis 5. Oktober 2026 (Ankündigung; vollständige Gewichte für Oktober 2026 angekündigt) · 5× · vu le 09 oct. 2026
Beam is Reflection AI's first open-weight language model, developed by a Brooklyn/New York startup founded in 2024 by former Google DeepMind researchers. It is a text-only sparse Mixture-of-Experts model with 501 billion total parameters, of which only 23 billion are active per token, pretrained on 23.8 trillion tokens. Beam targets coding, reasoning, and agentic workloads, and the company claims it matches Chinese open models like GLM-5.2 while using significantly less inference compute. Model weights were not yet released at announcement; a beta API via waitlist exists, but no official pricing has been published yet.
Fonctionnalités
| Key Benchmark (%) | Terminal-Bench v2.1: 80.1 (vs. GLM-5.2: 81.0, DeepSeek V4.1 Flash: 90.6, Kimi K3: 88.3) – vendor-reported |
| Context Window (Tokens) | 1 million tokens (effective, via midtraining); beta API currently limited to 256,000 tokens |
| License | Apache 2.0 (for weights, planned for October 2026) |
| Multimodality | No – text-only model, no image/audio input |
| Platform | Beta API (OpenAI-compatible, Chat Completions) via waitlist at platform.reflection.ai; weights not yet publicly available |
| Price per 1M Tokens | Not published; no rate card available in documentation or blog (as of October 6, 2026) |
| Release Date | October 5, 2026 (announcement); full weights planned later in October 2026 |