Enntity/d3-ultra-NVFP4

d3-ultra NVFP4 (for SparkScore)

d3-ultra is a 27B typed-decision model by the vLLM Semantic Router team. This repository holds it quantized to NVFP4, in the format NVIDIA ModelOpt uses for Qwen3.8-27B. It is what SparkScore serves on one NVIDIA DGX Spark with the Atlas engine: JevBench public 90.9% at a p50 of 146 ms per question.

You don't need to download this by hand. SparkScore's ./start.sh fetches it at a pinned revision and checks every file against install/d3-ultra-nvfp4.sha256.

What it is

• Source: vllm-sr/d3-ultra at revision ccdcfa611649a59e095e0b5e1732a15cedafadd6 (BF16, Apache-2.0).

• Format reference: a2genesis/Qwen3.8-27B-NVFP4 at revision bfd9b31207712e0850eec9da32261e8c5ee16af7 (ModelOpt MIXED_PRECISION).

• Quantization:

  • MLP weights: NVFP4 W4A16, 16-value blocks, with E4M3 block scales and an FP32 tensor scale.
  • Attention and Gated DeltaNet projections: FP8 per-tensor.
  • Everything else: unchanged from the source.

  The weight quantization reproduces ModelOpt's "max" calibration bit for bit. The FP8 activation scales (input_scale) are copied from the reference.

• Head: lm_head is built from d3-ultra's code readout (readout.safetensors). Its rows sit at the decision code token ids (decision_config.json), cast to BF16, and every other row is zero. Serve it with a BF16 LM head.

• MTP: d3-ultra has no MTP head, so none is included.

• Built with: SparkScore's tools/nvfp4_convert.py convert and validate, run by install/convert.sh. The conversion is deterministic: a fresh conversion matches these files byte for byte.

The model is meant for prefill-only scoring with prefix-bidirectional attention, which is how d3-ultra was trained. It is not a chat model. Use it through SparkScore's /v1/systemone API, or through Atlas's /v1/score with "attention": "prefix_bidir".

Results on one DGX Spark

JevBench v1.6.1, 231 public items, latency per question:

	Accuracy	Brier	p50	p95
This checkpoint, SparkScore (W4A4)	90.9%	0.118	146 ms	1.05 s
This checkpoint, SparkScore (W4A8)	92.2%	0.115	158 ms	1.31 s
d3-ultra BF16, d3's own runtime	91.3%	0.117	633 ms	3.07 s

Receipts and how to reproduce them are in the SparkScore repository (results/, docs/REPRODUCE.md).

License and credits

The weights are a quantization of d3-ultra and keep its license, Apache-2.0 (LICENSE). NOTICE is carried over from the source. d3-ultra is Copyright 2026 the vLLM Semantic Router Team. The quantization format follows NVIDIA ModelOpt; no ModelOpt code was used. Please cite d3-ultra when you use this model.