Published · Phase 3
What Can You Run With 128GB Unified Memory?
A 128 GB AI system can run 70B at high precision, 120B around 4–8 bit, and selected larger MoE checkpoints. The best use is capacity-heavy local work—not assuming every model will be fast simply because it fits.
Editorial review: complete · Updated 2026-08-30
Direct answer
Direct answer guidance: Choose 128 GB when a validated workload exceeds 80–96 GB or when local capacity/low power matters more than maximum token speed.
Model / quant variants
Model / quant variants guidance: llama.cpp offers the widest low-bit model access; vLLM/NIM/SGLang provides stronger serving when the model/architecture is supported.
Approximate weight footprint
Approximate weight footprint guidance: Use approximately 100–105 GB as a comfortable complete-workload ceiling. That leaves room for cache, runtime, OS, multimodal components, and modest concurrency.
Hardware tiers
Hardware tiers guidance: Use approximately 100–105 GB as a comfortable complete-workload ceiling. That leaves room for cache, runtime, OS, multimodal components, and modest concurrency. RTX PRO 6000 is faster below 85 GB; multi-GPU servers supply more bandwidth/concurrency; cloud GPUs handle intermittent giant-model work.
Observed configurations
Observed configurations guidance: Expect a broad range: 25–60 tok/s for many 27B–120B optimized workloads on Spark, lower for giant dense/MoE deployments, and much higher aggregate rates under batching.
Context / concurrency effects
Context / concurrency effects guidance: Use surplus memory for reliability first, then longer context or additional sequences. Advertised maximum context can still have impractical TTFT.
Performance expectations
Performance expectations guidance: Expect a broad range: 25–60 tok/s for many 27B–120B optimized workloads on Spark, lower for giant dense/MoE deployments, and much higher aggregate rates under batching.
Best-value configurations
Best-value configurations guidance: Capacity can avoid a server, but buying 128 GB for a 20 GB model wastes the platform's main advantage. Choose 128 GB when a validated workload exceeds 80–96 GB or when local capacity/low power matters more than maximum token speed.
What will not fit / weak evidence
- 128 GB unified memory is not equivalent to a 128 GB high-bandwidth discrete GPU.
- Reserve at least 15% rather than designing to the last gigabyte.
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.
- NVIDIA DGX Spark specifications
- NVIDIA RTX PRO 6000 Blackwell specifications
- OpenAI gpt-oss introduction and deployment guidance
- Mistral Small 4 model card
- DeepSeek-V4-Flash model card
- vLLM installation and hardware documentation
- llama.cpp project and backend documentation
- Qwen3.6-27B-NVFP4 vLLM Performance Report