AI model finder
Tell us the job.
We'll name the model.
Pick what you actually want to do and we rank the famous open models — Qwen, DeepSeek, Llama, Mistral and more — on the four things that decide it: speed, VRAM, accuracy for your task, and how open the license really is.
Task-fit accuracy
Benchmark-informed rating per job, not one leaderboard.
Real footprint
VRAM & tok/s from the engine at Q4_K_M.
Honest licenses
Apache/MIT vs community vs restricted.
Balanced weighs accuracy, speed, VRAM and license together.
Ranked for coding
—VRAM and tok/s are engine estimates at Q4_K_M on a single RTX 4090. Accuracy is a benchmark-informed editorial score per task — treat it as a strong starting shortlist, then verify on your own eval.
How we rank
Four axes, no hype.
Accuracy
A per-task score for each model, informed by public coding, chat, reasoning and long-context benchmarks — chosen for the job you picked, not a single overall number that hides task gaps.
Speed & VRAM
Straight from the engine: VRAM is the Q4_K_M weight footprint plus overhead; speed is the memory-bandwidth decode ceiling on the model's active params — so MoE models score fast even when they're large.
License
Apache 2.0 / MIT rank as fully open. Llama and Gemma are community licenses — free with conditions. Codestral (non-production) and OpenRAIL code models are restricted. Your priority can filter to open-only.
Picked a model?
Now see what it takes to run it.
Every recommendation has a VRAM number. The Workspace turns that into a GPU shortlist, a cost, and a buy-vs-rent call.
Evidence & method
How this calculation works
Filters the model catalogue by the VRAM your GPU can actually use, so you see only models that fit at a workable quantization.
Data sources
- NVIDIA, AMD & Intel GPU documentation
- Hugging Face model cards
- Official model papers
- Bitpute Methodology
Assumptions
- Weights = parameters × bits-per-weight ÷ 8
- Runtime overhead = 0.75 GB + 5% of weights (CUDA context, allocator, buffers)
- Recommended VRAM = total × 1.10 (10% headroom)
- Usable VRAM = card capacity × 0.95
- KV cache counted only when the model architecture is known; when applied, batch size 1, FP16 KV, context as entered
- Single GPU, inference workload
Limitations
- Actual VRAM varies by framework (PyTorch, llama.cpp, vLLM, TGI), driver/CUDA version and OS
- Quantization implementations differ; real bits-per-weight can vary from the nominal value
- When architecture is unknown the KV cache is omitted, so long-context use will exceed the estimate
- Estimates are for planning and comparison, not a guarantee
Data status: Hardware specs — vendor documentation. Model metadata — community model cards, not independently verified.
Related
Why does this estimate differ from other calculators?
- Different footprint formulas and rounding
- Different KV-cache assumptions (batch, context, precision)
- Framework and inference-engine differences
- Reserved and fragmented VRAM
- Driver and CUDA overhead
- Precision and quantization choices