DeepSeek-V4.1-Flash-1017 local setup comparison: ~75 tok/s on dual DGX Spark
A roundup of quantized local setups on dual DGX Spark, dual Strix Halo and a single Gorgon Halo 192GB, with Mia TensorFold decoding ~75 tok/s prose and 90–110 tok/s code.
DeepSeek-V4.1-Flash-1017 Local Setup Comparison (Oct 2026) 2x DGX Spark — Mia TensorFold Quant: EXL3-EAQv1 2.9 bpw Decode: ~75 tok/s prose / 90–110 tok/s code Prefill: 2000–2175 tok/s Concurrent: ~130 tok/s total at 4 streams Quality: High (calibrated, close to original) 2x DGX Spark — styles01/jaita tuned vLLM Quant: EXL3 2.9 bpw Decode: ~52 tok/s prose / ~94 tok/s code Prefill: Strong Concurrent: Excellent on structured (~127 tok/s) Quality: High 2x Strix Halo — Lucebox Quant: ROCMFP (2–3.5 bit experts) Decode: ~25–40 tok/s (est.) Prefill: Moderate Concurrent: Limited but usable Quality: Good (day-0 high-quality quants) Note: Dual-node claimed up to ~103 tok/s on code 2x Strix Halo — Project Maya (Maya-S) Quant: Custom 2–3 bit experts + 8-bit dense Decode: ~15–25 tok/s (est.) Prefill: Moderate Concurrent: Supported Quality: ~92.6% next-token match 2x Strix Halo — DwarfStar (dcapitella fork) Quant: Q2 GGUF (~2 bit) Decode: Lower Prefill: Moderate Concurrent: Limited Quality: Acceptable (more token-efficient) 1x Gorgon Halo 192GB — Lucebox (ds41-gorgon) Quant: ROCMFP2S (2-bit experts) Decode: ~30–50 tok/s code Prefill: Better than 128 GB Concurrent: Usable single-box Quality: Good Note: Best single AMD box 1x Gorgon Halo — Project Maya / DwarfStar Quant: Maya-S or Q2 Decode: Lower than Lucebox Prefill: Moderate Concurrent: Limited Quality: ~92% range