Bespoke Nimble
#20 en Modèles de langage open sourceUnknown · 3× · vu le 20 sept. 2026
36
Momentum
Bespoke Nimble (Bespoke-Nimble-9B) is an Open Source decision model released by Bespoke Labs, designed as an open reconstruction of "open Jev". It is a LoRA adapter based on Qwen3.5-9B, Fine-Tuned for context-based selection decisions, Booleans, and rubric ratings. The model evaluates permitted response Tokens directly via Logits/Softmax instead of generating text, enabling very fast Inference (approximately 100-115ms on H100). It is available under Apache-2.0 license on Hugging Face and GitHub and requires the Qwen3.5-9B base checkpoint.
Historique du momentum
22.06.20.09.
Fonctionnalités
| Key Benchmark (%) | 90.1% (90.12%) on curated 324-example eval, vs. 66.4% base Qwen3.5-9B and 93.2% Jev |
| Context Window (Tokens) | 2,048 tokens (training prompt limit) |
| License | Apache 2.0 |
| Multimodality | Not multimodal – pure text/decision classifier (no image/audio support documented) |
| Platform | Hugging Face / GitHub, runs on NVIDIA GPU (CUDA/BF16) or Mac with Apple Silicon |
| Price per 1M Tokens | $0.60 per 1M tokens (H100 inference, community benchmark) |
| Release Date | September 18, 2026 |