Tools

Find Your Setup

Pick your use case. We suggest a realistic model, the VRAM you need, the proposed card and the software — all with linked resources and reference data.

How it works Deterministic calculations grounded in published specs and real benchmarks. Every figure links to its source.

Choosing hardware and model

Before the math, read how to pick the right hardware and the model suited to your use: VRAM, quantization and response speed. Open the guide →

ResultLlama 3.1 8B · Q4_K_M· NVIDIA RTX 3060 12GBView details ↓

NVIDIA RTX 3060 12GB· 7.8 GB VRAM

Activity

Parallel users

Each person using the model at the same time adds memory: pick the maximum number of concurrent users.

Intelligence level

How much reasoning power do you need?

Context

How much conversation memory do you need?

Speed priority

Fits wellGeneral chat · Context: Medium

For General chat with 1 user: Llama 3.1 8B at 4-bit (Q4_K_M), using 7.8 GB of VRAM.

Suggested model

Llama 3.1 8B

4-bit (Q4_K_M) · 8B

Size chosen for General chat and Standard level: enough capability for your use, light enough to run smoothly on your hardware.

VRAM needed

Weights4.8 GB

The model's weights: the space it occupies in memory at rest.

KV cache2 GB

The conversation memory: it grows with context length and with the number of users.

Headroom1 GB

Safety space for the system and usage spikes.

Total7.8 GB

Computed for 16,384 tokens of context and 1 user. More context or users = more KV cache.

Proposed hardware

NVIDIA RTX 3060 12GB

12 GB VRAM · 360 GB/s

You need 7.8 GB of VRAM: the NVIDIA RTX 3060 12GB is the cheapest card with 12 GB, so it fits with headroom for 1 user.

Estimated speed

29.8

tokens per second (per user)

Fluid: responses stream while you read them.

Real-time simulator42 chars/s

▌

The text scrolls at the estimated speed: 1 token ≈ 1.4 characters.

Software needed

Ollama· Inference engine
Open WebUI· Web UI

Indicative budget

≈ CHF 1,104 – 1,868

GPU + base system

Alternative similar models

Other models with characteristics close to Llama 3.1 8B. Tap one to compare: the plan and the chart recompute instantly, and tapping it again reverts to the recommended model.

What if users increase?

1 👤 · NVIDIA RTX 3060 12GB · 7.8 GB2 👥 · NVIDIA RTX 3060 12GB · 9.8 GB3 👥 · NVIDIA RTX 3060 12GB · 11.8 GB4 👥 · NVIDIA RTX 4060 Ti 16GB · 13.8 GB5 👥 · NVIDIA RTX 4060 Ti 16GB · 15.8 GB

Each parallel user adds their own KV cache: NVIDIA RTX 3060 12GB · 7.8 GB for 1 user, NVIDIA RTX 4060 Ti 16GB · 15.8 GB for 5 users.

System specifications

Minimum and recommended for the whole system (GPU excluded — it's above). Indicative 2025/26 values for a desktop build; on a Mac the machine is the system.

	Minimum	Recommended
GPU / VRAM	NVIDIA RTX 3060 12GB · 12 GB VRAM
CPU	8 cores / 16 threads (e.g. Ryzen 5 / Core i5)	12+ cores / 20+ threads (e.g. Ryzen 7 / Core i7)
System RAM	16 GB	32 GB
Power supply	650 W	650 W
Storage	1 TB NVMe SSD	1 TB NVMe SSD
Cooling	Standard cooler for a single GPU	Standard cooler for a single GPU
Complete system (indicative)	≈ CHF 1,104 – 1,868

Model weights live in VRAM; system RAM serves the OS, loading and context.

Investment and VRAM

Model in the chart

How many parallel users fit on each card with the current model. Green bars reach your target; pick another model to see what changes.

Your current selection Your target (≥ 1 user) Below targetBar height is the estimated total price (card + system), in local currency.

Current setup: about 8 GB needed for 1 user.

Cloud alternative

If you prefer not to buy hardware, inference services run the same model on pay-per-use.

No particular vendor: the choice depends on budget, privacy and model type. The catalog lists the inference services, with details for each one.

Browse inference services

Sources and reference data

The result figures come from official specs and public benchmarks: each item below links directly to the underlying data.

Model:huggingface.co
Card:www.nvidia.com
Memory bandwidth and decode speed:www.nvidia.com

The tokens/s estimates derive from the card's memory bandwidth and published decode benchmarks.

Next step: install and use the setup

Pick the software you need and follow its resource page: you'll find the step-by-step install guide.

Ollama Open WebUI

Share this setup:/en/strumenti/setup

Saved configurations

Saved configurations stay in your profile: reopen, compare, delete.

Create an account Log in

HardwareIntermediate7 min

Hardware for Local AI: What You Need in 2026

The rule (weights + KV cache), real October 2026 prices for NVIDIA GPUs, Apple Silicon and CPU, the 36-month cost math and the choice by budget.

iaLocale Team - June 10, 2026Read

FundamentalsIntermediate6 min

How to Choose Your First Local AI Model

The three criteria for picking your first model: use case, memory (with the exact formula) and context. Recommended models for 8, 16 and 24 GB, and your first model in two commands.

iaLocale Team - June 9, 2026Read

ComparisonsIntermediate5 min

Ollama vs LM Studio: Which Tool to Choose

Ollama for machines and servers (API 11434), LM Studio for exploring and measuring (server 1234, built-in benchmark): the comparison table, the OpenAI-compatible API and the real case of using both together.