Qwen3.8-Flash-Next-Slipstream

Qwen3.8-Flash-Next for the Slipstream inference engine: our own NVFP4 quantization of the official BF16 checkpoint (routed experts and n-gram tables in NVFP4 with calibrated W4A4 activation scales, the rest kept in higher precision, MXFP8 prefill copies), plus the vision tower. Runs on one Spark (TP1) or two (TP2).

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: Qwen/Qwen3.8-Flash-Next. 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
qwen38fnc	model weights for one Spark (TP1)	single Spark
qwen38fnc_tp2_r0	model weights, rank 0 (TP2)	first
qwen38fnc_tp2_r1	model weights, rank 1 (TP2)	second
qwen38_vision	vision tower (BF16)	first

Download

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

hf download Interactify/Qwen3.8-Flash-Next-Slipstream --local-dir ~/vmodels \
  --include "qwen38fnc_tp2_r0/*" "qwen38_vision/*"

On the second Spark (rank 1):

hf download Interactify/Qwen3.8-Flash-Next-Slipstream --local-dir ~/vmodels --include "qwen38fnc_tp2_r1/*"

On a single Spark, download qwen38fnc (and qwen38_vision) instead:

hf download Interactify/Qwen3.8-Flash-Next-Slipstream --local-dir ~/vmodels --include "qwen38fnc/*" "qwen38_vision/*"

The tokenizer, chat template and generation config:

hf download Qwen/Qwen3.8-Flash-Next --exclude "*.safetensors"   # on the first Spark

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

Run

python server/slipstream.py serve qwen3.8-flash-next --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

• Prefill: 2110 (0.9k) / 3160 (3.5k) / 3410 (13.8k) tokens/s.
• Attention projections in 8-bit (standard): 4-6 % better perplexity than all-4-bit, for about 8 % more weight bytes.
• tool-eval-bench: 88/100, median agent turn 2.5 s (92 scenarios).

Notes

• Context up to 262,144 tokens.
• MTP drafts from the model's own MTP head.

License

Qwen Community License 1.0, as the base model. Read it before commercial use: it has conditions for very large products and for Model-as-a-Service / AI work assistant businesses. See LICENSE and NOTICE.