Instructions to use vllm-sr/d3-flash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use vllm-sr/d3-flash with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("zero-shot-classification", model="vllm-sr/d3-flash", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModel processor = AutoProcessor.from_pretrained("vllm-sr/d3-flash", trust_remote_code=True) model = AutoModel.from_pretrained("vllm-sr/d3-flash", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
D3 Flash leads 18 local decision configurations on an RTX 2060 6 GB
D3 Flash is the highest-scoring local configuration in my System 1 decision benchmark: 995/1,342 correct (74.1%), with a 408 ms median. The comparison now covers 49 routes, including 18 local configurations, on the same 31-category text bank.
I deliberately used an older RTX 2060 with 6 GB VRAM to explore a more accessible setup that readers may already have. The local D3 run used Artemis IQ3_M, served through Ollaya 0.17.0 over the LAN, one request at a time. Loading and warm-up calls are excluded.
| Local configuration | Accuracy | Median | p95 |
|---|---|---|---|
| D3 Flash Artemis 9B | 74.1% | 408 ms | 701 ms |
| Perplexity Flash v1.2 mixed4 | 72.4% | 798 ms | 1,526 ms |
| H2O-Lightning 4B Q8_0 | 66.5% | 259 ms | 459 ms |
D3 scored 43/46 on if-then rules, 40/43 on planning, and 48/51 on explicit preference ordering. Sequential procedures scored 14/48, and spatial tracking 22/49. Three requests exceeded this local runtime's input batch limit and remain unsuccessful in the full-bank score.
The hosted d3-turbo run scored 1,106/1,342 (82.4%), with a 165 ms median and 298 ms p95—the third-lowest median among the 27 hosted decision routes. That run was measured October 20; the local runs October 25, 2026. These are different serving setups.
Full results, questions, category profiles and methods · GitHub repository
