Published · Phase 2
Qwen3.6 35B Hardware Requirements
Qwen3.6-35B-A3B runs well on a 24–32 GB GPU at 4 bit, but 48 GB is the more flexible tier for long context and 8-bit experiments. DGX Spark is useful for BF16, multiple services, or very large context—not required for basic deployment.
Editorial review: complete · Updated 2026-08-30
Direct answer
Direct answer guidance: Choose 32 GB for cost-effective 4 bit, 48 GB for operational headroom, and 80–128 GB for high precision or large concurrent caches.
Model / quant variants
Model / quant variants guidance: Use a current vLLM/SGLang release for the official MoE architecture. A GGUF/llama.cpp path is appropriate only after exact model support is mature.
Approximate weight footprint
Approximate weight footprint guidance: Estimate 20–24 GB for 4-bit/NVFP4, 38–45 GB for 8 bit, and 75–85 GB for BF16 plus packaging overhead. Add cache, multimodal encoder, and runtime workspaces separately.
Hardware tiers
Hardware tiers guidance: Estimate 20–24 GB for 4-bit/NVFP4, 38–45 GB for 8 bit, and 75–85 GB for BF16 plus packaging overhead. Add cache, multimodal encoder, and runtime workspaces separately. Qwen3.8-27B is simpler and smaller; Qwen3.5-35B-A3B has a longer deployment history; 14B models are faster and easier for routine agents.
Observed configurations
Observed configurations guidance: Only about 3B of 35B parameters are active per token, enabling responsive inference when kernels are optimized. Do not use active parameters to estimate resident memory.
Context / concurrency effects
Context / concurrency effects guidance: Start at 16K–32K despite the much larger advertised maximum. Extreme context needs dedicated cache planning and long-prefill tolerance.
Performance expectations
Performance expectations guidance: Only about 3B of 35B parameters are active per token, enabling responsive inference when kernels are optimized. Do not use active parameters to estimate resident memory.
Best-value configurations
Best-value configurations guidance: A 32 GB consumer card is sufficient for most 4-bit users. Pay for memory only when context, concurrency, or precision is part of the requirement. Choose 32 GB for cost-effective 4 bit, 48 GB for operational headroom, and 80–128 GB for high precision or large concurrent caches.
What will not fit / weak evidence
- A model that loads at 4 bit may still exceed 32 GB with vision and long context.
- Validate the exact runtime release/model revision pairing before procurement.
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.