---
type: "Article"
title: "Qwen3.8-27B EXL3 with vLLM on 4× DGX Spark"
description: "Qwen3.8-27B EXL3 with vLLM on 4× DGX Spark: recipe by fujitsupolycom with speed, condition and source for every number, weights and license."
resource: "https://www.contextstudios.ai/local-ai/fujitsupolycom--qwen3.8-27b-exl3-k5k6-tp4-4x-dgx-spark"
language: "en"
generated:
  by: "process:contextstudios-md/1"
  at: "2026-10-02T20:37:58.265Z"
status: "stable"
---

# Qwen3.8-27B EXL3 with vLLM on 4× DGX Spark

Qwen3.8-27B EXL3 with vLLM on 4× DGX Spark: 35.1 tok/s according to github.com (dataset as of Sep 29, 2026).

- Creator: [fujitsupolycom](https://www.contextstudios.ai/local-ai/creators/fujitsupolycom.md)
- Engine: vLLM
- Quantization: EXL3 K5/K6 gemischt, 21.59 GB (hf hydrated); FP8 KV (kv_cache_dtype fp8); EXL3-Prefill FP8, Tile 256
- Model family: Qwen3.8-27B
- Hardware: dgx-spark x4
- Context: 1048576
- Artificial Analysis Intelligence Index: 33.7 ([source](https://artificialanalysis.ai/models/qwen3-8-27b))

## All numbers

- 35.1 tok/s (Everyday; Everyday, 1 request, realistic prompt, with Qwen MTP3 probabilistic mit standard rejection (per recipe)) — [source](https://github.com/FujitsuPolycom/sparkring/blob/main/performance/records/qwen38-27b/normalized-tp4-1m-probmtp-temp1-20260823.md)
- decode_tps: 35.07 (1 request) — [source](https://github.com/FujitsuPolycom/sparkring/blob/main/performance/records/qwen38-27b/normalized-tp4-1m-probmtp-temp1-20260823.md)
- decode_tps: 48.46 (1 request) — [source](https://github.com/FujitsuPolycom/sparkring/blob/main/performance/records/qwen38-27b/normalized-tp4-1m-probmtp-temp1-20260823.md)
- prefill_tps: 1964 (1 request) — [source](https://github.com/FujitsuPolycom/sparkring/blob/main/performance/records/qwen38-27b/normalized-tp4-1m-probmtp-temp1-20260823.md)

## Weights

- [malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated) (main, apache-2.0)

[https://github.com/FujitsuPolycom/sparkring](https://github.com/FujitsuPolycom/sparkring)

## Related

- [Local AI on your hardware.](https://www.contextstudios.ai/local-ai.md)