Published · Phase 3
Qwen3.5 35B Hardware Requirements
Qwen3.5-35B-A3B needs 24–32 GB for a practical 4-bit deployment, 48 GB for comfortable 8 bit, and around 80 GB for BF16. The RTX 5090 is the recommended single-user 4-bit platform; higher-memory systems target precision and concurrency.
Editorial review: complete · Updated 2026-08-30
Direct answer
Direct answer guidance: Buy 32 GB for one 4-bit service, 48 GB for 8 bit/longer context, and 96–128 GB for high-precision research or multiple resident models.
Model / quant variants
Model / quant variants guidance: Use vLLM for serving and current MoE kernels; use llama.cpp for a trusted GGUF when local simplicity is the priority.
Approximate weight footprint
Approximate weight footprint guidance: The official repository is about 72 GB. Budget 20–24 GB at 4 bit, 38–45 GB at 8 bit, and 75–82 GB at BF16 before cache and runtime workspaces.
Hardware tiers
Hardware tiers guidance: The official repository is about 72 GB. Budget 20–24 GB at 4 bit, 38–45 GB at 8 bit, and 75–82 GB at BF16 before cache and runtime workspaces. Qwen3.6 is the newer coding choice; Qwen3.8 is dense and simpler; Mistral Small 4 offers a larger MoE on high-memory hardware.
Observed configurations
Observed configurations guidance: A tuned 4-bit 5090 deployment can land around 80–160 tok/s, while a capacity-first Spark will be slower but can hold higher precision and much larger caches.
Context / concurrency effects
Context / concurrency effects guidance: Use 8K–16K on 32 GB and 32K on 48 GB after profiling. Larger advertised windows remain specialist configurations.
Performance expectations
Performance expectations guidance: A tuned 4-bit 5090 deployment can land around 80–160 tok/s, while a capacity-first Spark will be slower but can hold higher precision and much larger caches.
Best-value configurations
Best-value configurations guidance: The 4-bit consumer tier is the best value. Higher precision should earn its hardware premium in a controlled task evaluation. Buy 32 GB for one 4-bit service, 48 GB for 8 bit/longer context, and 96–128 GB for high-precision research or multiple resident models.
What will not fit / weak evidence
- Reserve memory for the vision tower when using multimodal inputs.
- Community quantization quality and metadata vary; prefer traceable conversions.
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.