Jared Palmer · siet 20. September 2026 (Kev-0.8B, 4B, 9B; laut ai-tldr.dev und AI Weekly). Kev-4B: Model-Card-Eintrag 24. Sep 2026; Kev-9B: · 2× · tolest 13. Okt. 2026
26
Momentum
Kev is an open-source family of small decision models by Jared Palmer, built on Qwen3.5 base models in 0.8B, 4B and 9B sizes. The models answer typed questions (yes/no, multiple choice, rating) with calibrated probabilities in a single forward pass, without generating text. The license is Apache-2.0; weights, training code and an API compatible with TypeSafe's System One are published.
Momentum-Verloop
20.07.18.10.
Features
| Key Benchmark (%) | Kev-9B: 85.2% out-of-domain accuracy on locked test (Jev: 85.7% on development items) |
| Context Window (Tokens) | 8,192 (validated context, Kev-9B) |
| License | Apache-2.0 |
| Multimodality | Text only (input: text; output: decisions/probabilities) |
| Platform | Open source, self-hostable (local GPU or Mac, MLX); also hosted on OpenRouter (Kev-4B only) |
| Price per 1M Tokens | Kev-4B on OpenRouter: $0.042 input / $0 output (self-hosting free, compute costs only) |
| Release Date | September 20, 2026 (0.8B, 4B, 9B) |