Will a model fit on my GPU? (VRAM)

For local inference, memory goes to three things, and we compute each one separately from the model's own config.json (layers, attention heads, key-value heads, head size, parameter count):

  • Weights= parameters × effective bits per weight for the quantisation you pick. We use the real effective size of each GGUF quant type, not its nominal name: a "4-bit" file is about 4.8 bits per weight once block scales and higher-precision tensors are counted.
  • KV cache = 2 × layers × KV heads × head size × context length × 2 bytes. Using key-value heads (not attention heads) is what makes this right for grouped-query attention models.
  • Overhead: a fixed allowance of about 0.8 GB for the driver context, activations and the runtime.

How we checked it: the bits-per-weight table is validated against 40 real published quantised file sizes (Llama 3.1 8B, Llama 3.3 70B, Qwen2.5 7B, Mistral 7B and Gemma 2 9B, as listed on ollama.com). Median error: 0.7%. Known limits: the 2-bit Q2_K quant is over-estimated on large models (we err towards "won't fit"), and models with mixed attention (some Qwen3.5/3.6 and Gemma 4 layers use linear or sliding-window attention) are modelled by their full-attention layers, with the other layers' small fixed cost noted on the model's page.

Mixture-of-experts models need memory for every expert, so their VRAM uses the full parameter count even though each token only reads the active experts.

How fast will it run? (tokens per second)

Generating text is limited by memory bandwidth, so our speed figure is the card's memory bandwidth divided by the bytes read per token (the active weights), derated to 75%. It is an estimate, not a benchmark. Your runtime, batch size and kernels change the real number. Where we haven't confirmed a card's bandwidth from a second source, we show no speed at all rather than a number we can't stand behind. We don't run our own hardware tests; measured results quoted in articles are attributed to whoever measured them.

What does it cost to rent? (cloud GPU prices)

  • Prices come from the public pricing APIs of Vast.ai, Runpod, Lium (its public node list) and Novita AI (the public GPU list behind its GPU page), captured every hour. For Lium we count only nodes available right now that can be rented as a single GPU, and keep spot and on-demand apart. Novita's list has about five single-GPU products; a card shows only while Novita reports it in stock, and disappears from our pages the hour it sells out.
  • We show the lowest listed hourly price per GPU and the time it was captured. Spot and interruptible prices can change or disappear without notice.
  • Only four providers are tracked, and Novita AI's public list covers a handful of cards. Other clouds may be cheaper; we say so rather than imply full coverage. Lium is a decentralized marketplace and its spot nodes can be interrupted.
  • Vast.ai, Runpod and Novita AI rental links are referral links, marked on the page; Lium links are plain links with no referral. They never change the price you pay or which card we list first: the order is cheapest first.
  • Novita AI and Massed Compute links are also referral links (added Oct 29, 2026). Neither is in the hourly price table. Where Novita serves the same model, the cost-to-run page shows its per-token API price (input and output, per 1M tokens), copied from Novita's own public model price list and dated with the day we read it; it is not refreshed hourly.

How we check AI creator tools (Higgsfield and others)

  • Plans, prices, credit costs and limits are read from the tool's own pricing page, help center, changelog and terms, and every answer page says the date we read them. Where the tool shows a price quote without generating (for example Higgsfield's cost preflight), we use that quote and say so.
  • Prices can differ by region and by visitor. We say where we read them from, and we tell readers to check the price shown to them before paying.
  • When a vendor's own pages contradict each other, the answer page shows both and says which one we followed.
  • We don't claim hands-on testing we didn't do. Sample images are labelled with the tool, the date and whether they were edited.
  • Every answer has a “What we couldn't verify” section and, where it fits, the alternatives and who should skip the tool.
  • Some links to these tools are referral links, marked on the page. We may earn a commission if you subscribe; it never changes the answer. We never publish promo codes we couldn't verify. All AI tool answers live on the Higgsfield hub.

How articles are researched

  • Each guide lists the sources it was built from under Sources, preferring primary ones: model cards and configs, vendor spec sheets, official pricing pages, changelogs and papers.
  • Articles are published by the Aliteq Team. We use AI tools to help draft, then check figures against those primary sources before publishing, and we show when a page was last updated.
  • Hardware performance claims always name who measured them.
  • Corrections are made in place with an updated date. See the editorial policy.

Use the tools