Qwen3.6 35B on DGX Spark

Measured / sourced results

What the numbers mean

What the numbers mean guidance: A public cross-platform report shows high aggregate throughput on DGX Spark, but aggregate multi-request numbers are not personal-chat latency. Expect responsive generation and test single-request decode separately before capacity planning. Begin with the official NVFP4 checkpoint, 32K context, and vLLM. Raise context toward the model's 262K native limit only if TTFT and cache consumption stay acceptable.

What stands out

Best DGX configuration: official NVFP4, vLLM, 32K starting context.

Memory estimate: about 20–24 GB at 4 bit and 38–45 GB at 8 bit.

Best use: coding and agentic workloads that benefit from long context and tool use.

Evidence quality / source

Evidence quality / source guidance: Model architecture and context come from Qwen; capacity is an explicit weight-bit estimate; observed throughput remains source-reported and configuration-specific. The 3B active-parameter figure describes compute routing, not storage: the complete 35B weight set still has to reside somewhere.

What is unknown

• The 3B active-parameter figure describes compute routing, not storage: the complete 35B weight set still has to reside somewhere.
• Aggregate output tok/s from a server benchmark should not be presented as one user's decode speed.

Workload fit

Workload fit guidance: Recommended for local coding agents, tool use, and long-context repository work where Qwen3.6's agentic tuning matters more than a dense model's simpler runtime support. Plan for roughly 20–24 GB for 4-bit/NVFP4 weights, 38–45 GB at 8 bit, and 75–85 GB at BF16 after packaging overhead. All three can fit, but 4-bit leaves far more room for its very long context.

Alternatives

Alternatives guidance: Qwen3.8-27B is the simpler dense alternative; Qwen3.5-35B-A3B has a broader deployment history; RTX 5090 is the faster hardware option when the chosen 4-bit checkpoint fits.

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.

• Qwen3.6-35B-A3B model card
• NVIDIA DGX Spark specifications
• vLLM installation and hardware documentation
• NVIDIA Developer Forums Qwen3.6 NVFP4 cross-platform report