Qwen3.8-27B-NVFP4-RTX5090
The flagship: NVIDIA ModelOpt NVFP4 of Qwen3.8-27B, tuned for GeForce RTX 5090.
- Full 262K context on 32 GB
- 81.6 tok/s alone — 155.8 tok/s with the drafter
- Accuracy parity with Unsloth NVFP4
Gittensor Model Hub
Fast, practical checkpoints for single-GPU Blackwell — RTX 5090 and RTX PRO 6000 — quantized, benchmarked, and served by our own engine, SparkInfer.
Qwen3.8-27B-DSpark-NVFP4
Matched DSpark drafter, trained and quantized against this exact checkpoint.
- 1.91× decode, outputs byte-identical
- 1.41 GB — a quarter of the built-in MTP head
- +14.2% acceptance on agentic tool calling (v2)
SparkInfer
Blackwell-native, zero-dependency inference engine.
- Custom RTX 5090 runtime
- 2–3× faster than llama.cpp in many workloads
More
No-MTP variant (−0.85 GB) DSpark drafter · BF16 source Spark-Hermes-3.8-27B · weights soon