From October 1‑9 2026 I posted on X (formerly Twitter) about the open‑weight d1 decision‑model family, highlighting releases, performance, ecosystem details, and community collaborations.
- d1‑3B (text + vision) and d1‑omni‑600M (text + image or audio) were announced as open‑weight, locally runnable decision models.
- Latency on an RTX 5090: ~70 ms for text‑only inference, ~100 ms when image input is included.
- d1‑3B achieved performance comparable to or better than GPT‑6.1 Sol and Claude Opus 5.5 on 4 of 6 real‑world tasks, while being 19×‑200× cheaper.
These releases were motivated by a desire to democratize fast decision‑making, enable local deployment, and reduce reliance on costly large‑language‑model APIs.
The ecosystem now includes collaborations with OpenRouter, the OpenAI Decision API, and hardware partners such as NVIDIA (DGX, RTX workstations) and Jetson edge devices.
- In the jev Decision Index v0.3, d1‑3B placed #25 at 2.3× the cost of the jev baseline.
- Median latency from Japan (text‑only) ≈ 240 ms.
- Cost per million input tokens on OpenRouter is $0.04, with $0 output cost and a 65K token context window.
- Hugging Face now hosts a “decision‑model” tag with 1,300+ models.
- TermGrade provides 1,000 executable terminal environments, graded against 6 models with 36,000 trajectories.