Command Code · 2× · naposledy 07. 10. 2026
43
Momentum
Agr is an Open Source decision model by Command Code that answers typed questions based on a given state (text or JSON) and provides probabilities for each answer option instead of free text. It comes in two sizes: Agr (31B parameters, based on Google DeepMind's Gemma 4 31B IT) and the smaller Agr-flash (360M parameters, based on Hugging Face's SmolLM2-360M). Agr achieves 58.15 points on the Decision Index 0.2.1 across 38 Benchmarks and is designed for routing, tool selection, and approval decisions. Both models are available under Apache 2.0 license on Hugging Face/GitHub and are deployed locally or self-hosted (no public API pricing found).
Vývoj momenta
10.07.08.10.
Vlastnosti
| Key Benchmark (%) | 58.15 on Decision Index 0.2.1 (38 benchmarks) |
| Context Window (Tokens) | Agr: 16,384 (state) / 16,384 (question) / 32,768 (total request) tokens; Agr-flash: 6,144 / 2,048 / 12,288 tokens |
| License | Apache 2.0 (base models Gemma 4 31B IT and SmolLM2-360M also Apache 2.0) |
| Multimodality | Text only (input as text or JSON); no image/audio support documented |
| Platform | Open-weight model, self-hosted via Python lib/server (agr serve) or Hugging Face; Agr served on 96 GB GPUs |
| Release Date | October 2026 (announced in Command Code blog, near Oct 6, 2026) |