DeepSeek-V4.1-Flash-Slipstream

DeepSeek-V4.1-Flash (1017) for the Slipstream inference engine: our own EXL3 quantization of the routed experts with K2/K3/K4 bit widths per expert (3.04 bpw, chosen from routing statistics and calibrated on 1017), dense weights in MXFP8 (lossless from FP8), the checkpoint's DSpark block drafter, the Engram tables and the vision tower.

These are Slipstream weight files (weights.vbin + index.json per directory), the engine's own pre-packed format. They run only with Slipstream; they are not safetensors and do not load in other runtimes. Built with Slipstream's converters at commit 20c2dd9 (2026-10-31).

Base model: deepseek-ai/DeepSeek-V4.1-Flash-1017. Engine: Slipstream (engine code release coming soon). Hardware: two NVIDIA DGX Spark (GB10, 128 GB each) connected over their 200 GbE link.

Contents

Directory	What	Node
dense1017_r0	dense weights + Engram rows, rank 0	first
dense1017_r1	dense weights + Engram rows, rank 1	second
x3_1017_r0	EXL3 routed experts, rank 0	first
x3_1017_r1	EXL3 routed experts, rank 1	second
dsp_1017_r0 / _r1	DSpark drafter	first / second
head4_1017_r0 / _r1	4-bit output head for drafting	first / second
dx3_1017_r0 / _r1	4-bit dense copies for the fast profile	first / second
vision1017	vision tower	first
orig	tokenizer, config and chat template	first

Download

On the first Spark (rank 0, runs the HTTP server):

hf download Interactify/DeepSeek-V4.1-Flash-Slipstream --local-dir ~/dsv41 \
  --include "*_r0/*" "vision1017/*" "orig/*"

On the second Spark (rank 1):

hf download Interactify/DeepSeek-V4.1-Flash-Slipstream --local-dir ~/dsv41 --include "*_r1/*"

The tokenizer, config and chat template are included (orig/); nothing else needs downloading.

The directory names match the defaults in Slipstream's server/models.json, so no configuration is needed.

Run

python server/slipstream.py serve deepseek-v4.1-flash --profile balanced

The server speaks the OpenAI and Anthropic APIs on port 8080. See the Slipstream README for profiles, the router and engine settings.

Measured on Slipstream

• Decode, one stream (fast profile): structured 126 / prose 62 / code 97 / JSON 105 tokens/s; four streams in total 136 / 101 / 131 / 157.
• Prefill 8K-32K: about 2,300 tokens/s (exact).
• Quality against a CPU reference with the original MXFP4 experts (5 texts of 1024 tokens): KL 0.007-0.041, top-1 90-98 %, perplexity within 0.5 %.
• Greedy output with DSpark drafts is identical to decoding without drafts.

Notes

• Context up to 524,288 tokens.
• Needs both Sparks (no single-Spark variant).

License

MIT, as the base model (DeepSeek). See LICENSE and NOTICE.