| d3-ultra (Decision 3.0) | |
| Copyright 2026 vLLM Semantic Router Team | |
| Built on Qwen/Qwen3.5-27B (Apache-2.0). | |
| Modified by the FastFlowLM team: weights quantized to the Q4NX format (Q4_K projections, Q8_0 token embedding) for AMD NPUs and repackaged for FastFlowLM. The readout, decision config, tokenizer and chat template are unmodified. | |