Qwen3.8 27B on NVIDIA DGX Spark

Measured / sourced results

What the numbers mean

What the numbers mean guidance: Public and StackBench-adjacent DGX Spark runs cluster around 28–31 decode tok/s. A sensible planning band is 25–40 tok/s for one request, with prompt processing and aggregate throughput varying more widely. Start with 4-bit weights, 16K context, and concurrency one. Increase context or precision only after measuring the remaining unified-memory budget.

What stands out

Expected single-request decode: roughly 25–40 tok/s.

Recommended starting point: 4-bit weights, 16K context, concurrency one.

Choose RTX 5090 instead when speed matters and 4-bit quality is acceptable.

Evidence quality / source

Evidence quality / source guidance: The file size is from the official Qwen repository; the performance band is an estimate anchored by public DGX Spark runs near 30 tok/s. Quantization quality varies by quantizer and calibration; Q4_K_M or an equivalent high-quality 4-bit format is a starting point, not an absolute winner.

What is unknown

• Quantization quality varies by quantizer and calibration; Q4_K_M or an equivalent high-quality 4-bit format is a starting point, not an absolute winner.
• Do not compare aggregate server throughput with single-request decode speed.

Workload fit

Workload fit guidance: Recommended for private coding, document work, multimodal experimentation, and local agents where 25–40 tok/s is interactive enough and long context matters more than minimum latency. The official BF16 repository is about 55.6 GB. Expect roughly 16–19 GB for a practical 4-bit GGUF and about 29–33 GB at 8 bit, plus KV cache and runtime overhead—well inside DGX Spark's working budget.

Alternatives

Alternatives guidance: RTX 5090 is substantially faster for this model because it fits in 32 GB at 4 bit. DGX Spark wins when you want 8-bit weights, much longer context, or multiple resident models.

Planner CTA

Compare this recommendation against your model, context, concurrency, latency, and budget in the ComputeSage Planner.

Sources / methodology

Recommendations combine official specifications, public model/runtime documentation, adjacent benchmark observations, and clearly labeled engineering estimates. Review the ComputeSage evidence methodology.

• Qwen3.8-27B model repository
• NVIDIA DGX Spark specifications
• vLLM installation and hardware documentation
• llama.cpp project and backend documentation
• StackBench controlled lab approved source qwen38-dgx-spark-owner-archive-20260817