Convly is the best place to compare AI models alongside the GPUs that run them. Our models database lists params, context, and live API prices side by side, while three free calculators estimate the VRAM each model needs, the API cost per month, and whether self-hosting beats paying per token — the model-plus-hardware view no single tool offered before.
Most comparison sites answer only half the question. Benchmark leaderboards rank models but never tell you what GPU they need; hardware sites list GPUs but never map them to a specific model at a specific quantization. This page puts both in one table, then hands you the tools to plug in your own numbers.
AI models paired with the GPU to run them
The table below pairs ~12 popular models with their approximate parameter count, the minimum VRAM to run them at 4-bit quantization, a recommended consumer GPU, and a rough API price tier. VRAM figures follow the rule of thumb of roughly 0.5–0.6 GB per billion parameters at 4-bit, plus 1–3 GB for the KV cache. An em-dash (—) means the model is API-only or too large to run practically on consumer hardware. All prices are pulled from our Database di modelli IA.
| Modello | Approx params | Min VRAM (4-bit) | Recommended consumer GPU | API price tier ($/1M in→out) |
|---|---|---|---|---|
| Mistral 7B | 7B | ~4–5 GB | RTX 4060 (8 GB) | Ultra-low ($0.02 → $0.03) |
| Llama 3.1 8B | 8B | ~5 GB | RTX 4060 (8 GB) | Ultra-low ($0.02 → $0.03) |
| Qwen3 14B | 14B | ~8–10 GB | RTX 3060 (12 GB) | Low ($0.12 → $0.24) |
| Gemma 3 27B | 27B | ~16–18 GB | RTX 4090 (24 GB) | Ultra-low ($0.08 → $0.16) |
| Qwen3 32B | 32 miliardi | ~20 GB | RTX 4090 (24 GB) | Low ($0.08 → $0.28) |
| Llama 3.3 70B | 70B | ~40–48 GB | RTX 6000 Ada (48 GB) or 2× RTX 4090 | Low ($0.10 → $0.32) |
| DeepSeek R1 Distill Llama 70B | 70B | ~40–48 GB | RTX 6000 Ada (48 GB) | Low-mid ($0.80 → $0.80) |
| DeepSeek V4-Flash | — (large MoE) | — | — (API-only in practice) | Ultra-low ($0.14 → $0.28) |
| Claude Haiku 4.5 | — | — | — (cloud) | Mid ($1.00 → $5.00) |
| Gemini 3.1 Pro | — | — | — (cloud) | Mid ($2.00 → $12.00) |
| Claude Opus 4.8 | — | — | — (cloud) | Premium ($5.00 → $25.00) |
| GPT-5.5 | — | — | — (cloud) | Premium ($5.00 → $30.00) |
Two patterns jump out. First, open-weight models scale down to hardware most people already own — a 7–8B model fits an 8 GB card, while a 32B model needs a single 24 GB RTX 4090. Second, closed frontier models can only be rented by the token, and the price gap is enormous: our Indice prezzo-prestazioni per l'IA measured a 114× blended-cost spread across the field, from about $0.18 to $20 per 1M tokens.
Three free tools to run your own numbers
The table gives you the shape of the trade-off. These three calculators let you pin down the exact answer for your model, your hardware, and your volume.
1. LLM VRAM Calculator
Pick a model size (or enter a custom parameter count), a quantization level, and a context length, and the LLM VRAM calculator tells you how much GPU memory the model needs and whether it fits a given card. It is the fastest way to check "will a 32B model run on my RTX 4090?" before you download 20 GB of weights.
2. AI API Cost Calculator
If you would rather rent than own, drop your monthly input and output token volume into the Calcolatore dei costi delle API per l'IA and it estimates the monthly bill for each model, using prices pulled live from our models database. It is the quickest way to see how much you save by moving from a premium model like GPT-5.5 to an ultra-low tier like DeepSeek V4-Flash.
3. Self-Hosting vs API Calculator
The buy-versus-rent decision comes down to volume. The calcolatore per il confronto tra self-hosting e utilizzo di API takes your token volume, GPU purchase price, amortization period, electricity rate, and utilization hours, then shows the break-even point where owning a GPU beats paying per token. Below roughly 50M tokens a month, per-token pricing usually wins; a well-utilised owned GPU only pulls ahead at high, steady volume.
Domande frequenti
How much VRAM do I need to run a 70B model?
At 4-bit quantization a 70B model needs roughly 40–48 GB of VRAM for the weights, plus another 1–3 GB for the KV cache. In practice that means a single 48 GB card such as an RTX 6000 Ada, or two 24 GB RTX 4090s in parallel. Running at 8-bit roughly doubles the requirement. Use the Calcolatore VRAM to check your exact context length.
È più economico eseguire in auto-hosting oppure utilizzare un’API?
Below roughly 50 million tokens per month, per-token API pricing almost always wins. A $2,000–$2,500 GPU amortized over three years, with electricity, runs about $85–$130 per month all-in, so it only pays off at high, steady volume — where a well-utilised owned GPU can cut inference cost by up to ~5×. Run your own volume through the calcolatore per il confronto tra self-hosting e utilizzo di API to find your break-even point.
Which GPU is best for local LLMs in 2026?
For most people the RTX 4090 (24 GB) is the sweet spot — it runs 32B models at 4-bit comfortably and handles long contexts. An 8 GB RTX 4060 covers 7–8B models, a 12 GB RTX 3060 runs 8B models with headroom, and stepping up to a 70B model requires a 48 GB card. Match your target model to a card with the Calcolatore VRAM.
What is the cheapest AI model per token?
DeepSeek V4-Flash is the cheapest in our database at $0.14 input and $0.28 output per 1M tokens (about $0.18 blended) — roughly 114× cheaper than the priciest frontier model and about 37× more intelligence per dollar than Claude Opus 4.8. See the full ranking in the Indice prezzo-prestazioni per l'IA.
