Qwen3.5 35B on DGX Spark

Measured / sourced results

What the numbers mean

What the numbers mean guidance: Public reports range widely by checkpoint and runtime. On DGX Spark, plan on interactive single-user output in the tens of tok/s; sparse MoE activation can raise throughput, but do not reuse an aggregate server number as personal-chat speed. Start with a reviewed 4-bit checkpoint and 16K–32K context. Move to 8 bit only if your own task evaluation finds a quality gain worth the extra memory.

What stands out

Comfortable at 4, 8, or 16 bit on DGX Spark.

Use vLLM for a server; llama.cpp for flexible local quantization.

Keep the starting context modest and expand only after measuring cache headroom.

Evidence quality / source

Evidence quality / source guidance: Qwen supplies the model identity and repository size; fit ranges use transparent bit-per-weight math plus overhead. Check the exact model revision and runtime release notes before downloading a 70+ GB checkpoint.

What is unknown

• Check the exact model revision and runtime release notes before downloading a 70+ GB checkpoint.
• Vision inputs and large contexts add memory beyond the language-model weight estimate.

Workload fit

Workload fit guidance: Recommended for users who want a capable multimodal/coding MoE with enough memory for long context and local tools. A 32 GB GPU is still the speed/value option at 4 bit. The BF16 repository is about 72 GB. Budget roughly 20–24 GB at 4 bit, 38–45 GB at 8 bit, or around 75–82 GB at BF16, before context and runtime overhead.

Alternatives

Alternatives guidance: Qwen3.6-35B-A3B is the newer coding-focused choice; Qwen3.8-27B is a denser, simpler deployment; Mistral Small 4 provides a larger 119B MoE when 128 GB capacity is the priority.

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.

• Qwen3.5-35B-A3B model repository
• NVIDIA DGX Spark specifications
• vLLM installation and hardware documentation
• llama.cpp project and backend documentation
• NVIDIA Developer Forums Qwen3.5 SM121 kernel benchmark