DeepSeek-V4.1-Flash-1017 · Strata (Strata-DS fork) · 1× DGX Spark (GB10, 128GB unified memory, 4TB NVMe)
English | 简体中文 | 日本語 · ← All recipes
Source level: ✅ Measured · Date: 2026-10-18 · Numbers come from the source. This repo did not measure them again.
Hardware
- 1× DGX Spark (GB10, 128GB unified memory, 4TB NVMe) (DGX Spark)
- Reference cost: ≈ $6,950
Model
- DeepSeek-V4.1-Flash-1017 ·
LibertAIDAI EXL3 2.9bpw (experts; dense 5 / head 6) + the official 1017 checkpoint's other tensors and Engram (single machine, DSpark drafting every step, SSD tail skip) - Download: huggingface.co/LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw · huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
Engine
- Strata (Strata-DS fork) — install: https://github.com/Jackwwg83/Strata-DS
- Note: Strata-DS (Strata fork with its own DeepSeek engine ds41_serve, C++/CUDA; hot experts in a unified-memory RAM tier computed in place by the GPU, cold experts read from the SSD on demand; sudoingX's exl3_smoe and exl3_mrow EXL3 kernels)
Launch command
git clone https://github.com/Jackwwg83/Strata-DS.git && cd Strata-DS && ./setup.sh --setup --yes --family deepseek --no-start && ./run-deepseek-1017-exl3-2.9bpw.sh
Key parameters
- Context: 256K (262,144)
- Concurrency: Single stream
- Tags:
2bit,MoE,long-context,speculative,offload
Expected speed
- Real coding-agent session, 48 turns (context 10K–91K), 70,211 tokens at 30.8 tok/s average (per turn 23.6–36.4, median 28.6); single runs: decode at 8K/32K/64K/128K/250K context 24.2/27.4/27.4/27.3/22.9; time to first token for an agent turn adding ≤50 tokens median 0.57 s, a new conversation with the same 10K system prompt 0.25 s; cold prefill about 500 tok/s (250K in 8.3 min); vs official MXFP4 KL 0.074 / top-1 88.6% (✅ Measured)
Caveats
- Speed changes with drivers, engine version, context length, and prompt length.
Source & attribution
- https://github.com/Jackwwg83/Strata-DS/blob/main/docs/DEEPSEEK_SPARK.md
- Credit to the original author. This page only summarizes and links. Data under CC BY 4.0.
Source file: recipes/comm-056.md