The amount of RAM you have is the single biggest constraint on which local AI models you can run. A 7B parameter model at Q4 quantisation fits in 16GB with room to spare. A 70B model at Q4 needs around 40GB, requiring 64GB RAM to run comfortably. Understanding which models fit at each memory tier stops you from downloading a model that will either refuse to load or crawl at unusable speeds.
In short: Model fit depends on the actual model artifact, quantisation format, architecture, context length, cache precision, concurrency, and runtime overhead; parameter count alone does not define each RAM tier. Beyond 64GB includes high-end personal workstations and unified-memory desktops as well as enterprise and multi-GPU systems. NAS memory limits vary by model: some consumer units officially max out at 16GB or 32GB, while others support 64GB or more.
How RAM Requirements Are Calculated
The working formula is straightforward: RAM needed (GB) = (parameter count in billions) x (bits per weight) / 8 x 1.1. The 1.1 multiplier accounts for the key-value cache and runtime overhead.
A 7B model at Q4 (4 bits per weight): 7 x 4 / 8 x 1.1 = ~3.9GB for the model weights. Add 2-4GB for the context cache and OS, and you need roughly 6-8GB free RAM to run it. A 13B model at Q4 needs approximately 8-10GB. A 70B model at Q4 needs roughly 38-42GB.
This is why quantisation level matters so much. For the same model, Q8 weights generally require substantially more memory than Q4 weights, but the total runtime requirement does not increase by an exact factor of two. Higher quantisation means better output quality but higher memory cost.
16GB RAM: What Actually Works
16GB is the entry tier for genuinely useful local AI in 2026. It is the RAM ceiling on most consumer NAS devices, a common default on mid-range mini-PCs, and still the configuration on a large number of laptops and older desktops.
At 16GB, a Q4_K_M 7B or 9B model fits with room for OS overhead and context, producing output that is practically useful for document Q&A, summarisation, and basic code completion. CPU inference speed varies substantially with the exact processor, memory bandwidth, model, quantisation, backend, thread settings, context length, and concurrent NAS workload; benchmark the intended configuration before treating it as suitable for interactive use. Functional for batch tasks, noticeably slower for interactive chat.
16GB RAM: What Runs and How Well
| Model | Quantisation | Approx RAM Used | Quality vs GPT-4 | |
|---|---|---|---|---|
| Llama 3.1 8B | Q4_K_M | ~5.5GB | ~5.5GB | Good for factual Q&A, coding basics |
| Mistral 7B | Q4_K_M | ~4.8GB | ~4.8GB | Strong reasoning, fast |
| Gemma 2 9B | Q4_K_M | ~6.2GB | ~6.2GB | Solid general-purpose |
| Phi-3 Mini 3.8B | Q4_K_M | ~2.5GB | ~2.5GB | Fast, good for simple tasks |
| Llama 3.1 8B | Q8_0 | ~9.0GB | ~9.0GB | Better quality, needs 12GB free |
| Llama 3 70B | Q2_K | ~28GB | ~28GB | Does NOT fit in 16GB |
At 16GB, a Q4_K_M 7B-9B model leaves enough headroom for the OS, Ollama's server process, and a browser or two. , and it works adequately for document summarisation, Q&A on local files, and basic code completion.
The ceiling at 16GB is quality. A 7B model at Q4, regardless of family, will not match GPT-4o or Claude Sonnet on complex reasoning, nuanced writing, or multi-step analysis. That gap is not simply a configuration problem; capability also depends on model architecture, training, post-training, task specialisation, and quantisation.
At 16GB, some 12B-13B Q4 models can fit with a moderate context, but they leave less headroom for the OS, concurrent services, and long contexts., run multiple models simultaneously, or maintain long context windows (32K+ tokens) without degrading performance.
32GB RAM: The Practical Sweet Spot
32GB opens up the 13B-14B model tier, which represents a meaningful quality jump over 7B. Models like Qwen 2.5 14B and Mistral NeMo 12B run comfortably at Q4-Q6 quantisation on many 32GB systems, subject to context and runtime overhead.
More importantly, 32GB lets you run a 7B model at Q8 (near-lossless quality) while still having RAM headroom for context and concurrent processes. This is a better use of 32GB than forcing a 13B model at Q3, which introduces quantisation artefacts that degrade output consistency.
32GB RAM: Practical Configurations
| Model | Quantisation | Approx RAM Used | Notes | |
|---|---|---|---|---|
| Llama 3.1 8B | Q8_0 | ~9.0GB | ~9.0GB | Best-quality 7B, plenty of headroom |
| 13B-14B dense model | Q4_K_M | Varies by model | Varies by model | Check the exact model artifact |
| Qwen 2.5 14B | Q6_K | ~12.5GB | ~12.5GB | Strong coding and Chinese language |
| Mistral Nemo 12B | Q4_K_M | ~8.0GB | ~8.0GB | 128K context window, efficient |
| DeepSeek Coder 33B | Q3_K_M | ~16GB | ~16GB | Squeezes into 32GB, quality compromise |
| Llama 3 70B | Q2_K | ~28GB | ~28GB | Technically fits but very poor quality |
The notable mistake at 32GB is trying to run a 70B model with extreme quantisation (Q2). Q2 quantisation causes material quality loss; compare the specific large Q2 and smaller Q4 models on representative workloads instead of assuming either will always be better. The numbers fit but the results disappoint. If 70B quality is the target, 64GB is the right starting point.
On a NAS with 32GB RAM, the CPU becomes the bottleneck before RAM does. The DS925+ uses an AMD Ryzen V1500B, not a J4125 or R1600; inference throughput should be compared using benchmarks of the exact NAS and mini-PC configurations.
64GB RAM: 70B Models Become Practical
64GB is the threshold where running a 70B model at Q4 becomes practical. Ollama's Llama 3.1 70B Q4_K_M artifact is approximately 43GB, leaving about 21GB of a nominal 64GB system before accounting for the OS, context cache, compute graph, and other runtime overhead. This produces quality notably closer to GPT-4o than a 7B or 13B model, particularly on complex reasoning and long-form generation tasks.
64GB also enables multi-model setups: running two 7B models simultaneously for specialised agent pipelines (one for planning, one for execution), or loading a 13B model alongside a smaller specialised model for coding or summarisation.
64GB RAM: What Opens Up
| Model | Quantisation | Approx RAM Used | Notes | |
|---|---|---|---|---|
| Llama 3.1 70B | Q4_K_M | ~40GB | ~40GB | Practical 70B quality, recommended config |
| Llama 3.1 70B | Q6_K | ~57GB | ~57GB | Near-lossless at 64GB, tight headroom |
| Qwen 2.5 72B | Q4_K_M | ~47GB before runtime overhead | ~47GB before runtime overhead | Strong multilingual + coding performance |
| Mixtral 8x7B MoE | Q4_K_M | ~28GB | ~28GB | MoE: fast inference for similar quality |
| Two x Llama 3.1 8B | Q4_K_M | ~12GB total | ~12GB total | Multi-agent, each model isolated |
At 64GB, CPU inference speed becomes the limiting factor more acutely. CPU-only 70B generation speed varies widely with the processor, memory bandwidth, model, quantisation, backend, context and thread settings; benchmark the exact system before treating it as practical for interactive use. This is usable for document processing and batch tasks, but is noticeably slow for interactive chat. Adding a GPU changes this equation entirely, but that requires hardware capable of GPU expansion.
128GB RAM: Higher-Precision 70B and Multi-Model Pipelines
128GB configurations are becoming accessible in 2026 through workstation-class mini-PCs and desktops with large DIMM slots. At 128GB, Q6 or Q8 versions of a 70B model can fit with more headroom than FP16, reducing quantisation loss relative to Q4; the practical quality difference is model- and task-dependent.
128GB also enables multi-model agent pipelines: running two large models simultaneously, one for planning and one for execution, or pairing a 70B general model with a smaller specialised coding model. This is primarily relevant for power users running AI as a primary workload rather than as a background service alongside NAS storage duties.
For most readers building a NAS with AI as a secondary capability, 32-64GB remains the practical sweet spot. 128GB starts to make sense when a user needs to run larger or less-quantised local models, longer contexts, or multiple models without relying on cloud inference.
NAS RAM Limits: What This Means for AI
Consumer NAS memory capacities vary by model: current examples range from 16GB and 32GB official maximums to 64GB or more. The QNAP TS-464 supports up to 16GB. The Synology DS925+ supports up to 32GB. UGREEN's DXP6800 Pro supports up to 64GB. QNAP's workstation TVS-H series supports 64GB+.
A NAS's practical AI ceiling must be determined from the exact model artifact, architecture, quantisation, context requirement, runtime overhead, and CPU performance rather than a fixed parameter tier. That is adequate for local document Q&A, summarisation, and simple automation. It is not adequate for quality comparable to current frontier models.
The implication: for users whose primary goal is private local AI inference at higher quality, a dedicated mini-PC with more RAM and a faster CPU will outperform a NAS running Ollama even with equivalent RAM installed. The NAS wins when storage capacity and AI inference are both required from a single appliance.
Common Mistakes to Avoid
Mistake 1: Downloading a 70B model onto a 16GB device. If the model and runtime allocations exceed available physical memory, the system may swap heavily, fail the allocation, or terminate the process depending on the OS and configuration; any successful swapped execution can be extremely slow. Always check model size against available RAM before downloading.
Mistake 2: Assuming Q2 quantisation of a large model beats a well-quantised small model. Q2 causes material quality loss; test the specific large Q2 and smaller Q4 models on representative tasks rather than assuming model size will outweigh quantisation loss. Match the model to your RAM tier rather than forcing an oversized model in.
Mistake 3: Ignoring context window RAM overhead. A 7B model at Q4 uses ~5GB for weights, but a 32K token context window adds another 2-4GB. Long context sessions will cause swapping on systems with limited headroom even if the model base fits.
Australian Context: RAM Upgrade Costs
NAS RAM upgrade costs in Australia are reasonable for most models. Australian DDR4 SO-DIMM prices are volatile and vary substantially by capacity, ECC support, vendor validation, and stock; check current prices for the exact compatible module. The TS-464 officially maxes out at 16GB. The DS925+ officially supports 32GB using two matched 16GB Synology ECC SO-DIMMs; check current pricing for those exact modules.
Note that some Synology models use proprietary memory configurations or have limitations on third-party RAM. The DS925+ uses DDR4 ECC SO-DIMM modules, but Synology specifies its own compatible module models and warns that non-Synology expansion memory can affect stability, warranty, and support. Earlier Synology models like the DS923+ had some compatibility constraints. Check the vendor's memory compatibility list, or use Crucial's compatibility tool for your model before purchasing.
The Beelink SER8 and Minisforum HX90G use replaceable SO-DIMM memory; other mini-PCs may have soldered RAM, so confirm the exact configuration before purchase. Confirm the configuration at purchase; upgrading after the fact may not be possible.
Related reading: our NAS buyer's guide, our NAS vs cloud storage comparison, and our NAS explainer.
Use our free AI Hardware Requirements Calculator to size the hardware you need to run AI locally.
For a full explanation of what Q4, Q6, and Q8 quantisation means for model quality and RAM requirements, see the LLM quantisation guide.
Can I run Llama 3 70B on 32GB RAM?
Technically yes at Q2_K quantisation, whose model artifact is about 26GB before runtime overhead. Q2 causes material quality loss, so compare the intended 70B Q2 model with suitable smaller models on representative tasks. If 70B quality is the goal, 64GB is the correct RAM target. If you have 32GB, use a 13B model at Q4-Q6 instead.
Does GPU VRAM work the same as system RAM for local AI?
GPU VRAM and system RAM serve different roles. When a model is loaded onto a GPU, it uses VRAM. When it runs on CPU only (as on most NAS), it uses system RAM. Many 7B Q4 models can fit fully in 8GB VRAM at a modest context, but KV cache, compute-graph, display, and runtime allocations must also fit. Full-GPU execution is generally preferable; hybrid and CPU-only performance depends on the GPU, CPU, memory bandwidth, interconnect, backend, and offload split. See our GPU expansion guide for NAS-specific GPU details.
What is the best model to run on a NAS with 16GB RAM?
Llama 3.1 8B at Q4_K_M or Q6_K is an older but compatible starting point for 16GB systems; compare it with newer models that fit the same memory, context, and workload requirements. Gemma 2 9B and Mistral 7B are strong alternatives. Models above 9B can also fit in 16GB when their actual artifact and runtime allocations leave sufficient headroom; check the exact model tag, context requirement, and loaded memory use.
How do I check how much RAM a model will use before downloading it?
On the Ollama model library page, each model variant lists the quantisation level and size. Stored model size is only a starting point. Add context-cache, compute-graph, runtime, concurrency, and OS requirements, which vary by model and configuration. After download, use `ollama list` to see the stored model size and `ollama ps` to inspect the loaded allocation, processor split, and context. For pre-download estimates, multiply billions of parameters by quantisation bits, divide by 8, and add 20%.
Does ECC RAM matter for local AI inference?
ECC RAM matters for data integrity in storage and server workloads where a bit flip could corrupt a database or file. ECC is not generally required to run local LLM inference, but it can improve system reliability by detecting or correcting memory errors; the effect of an uncorrected error is not necessarily limited to one token. ECC is not required for local LLM inference, though some QNAP workstation NAS models support it. Standard non-ECC DDR4 or DDR5 SO-DIMMs are fine for Ollama/LLM use.
Can I run local AI on a NAS with only 8GB RAM?
Yes, but the options are constrained. Phi-3 Mini (3.8B) and other sub-4B models run well on 8GB. Llama 3.2 3B is a capable general-purpose option. Standard 7B models at Q4 technically fit but leave very little headroom, and performance degrades noticeably with concurrent NAS operations. 8GB is workable for experimentation, but 16GB is the practical minimum for reliable daily use.
Want to know whether your specific NAS hardware can handle local AI workloads? The AI NAS hardware requirements guide covers CPU, RAM, NPU, and storage considerations in detail.