Can My PC Run a Local LLM? VRAM & RAM Calculator
Pick your GPU or enter your VRAM, system RAM, and target context length (8K–128K). Get an instant quantization verdict, KV-cache memory estimate, expected tok/s, and local cost vs API break-even.
Yes — your hardware can run Qwen3.8-Flash-Next locally.
Q2_0
130.1 tok/s
~4 GB (32K)
$0.102/M
1104 days
| Quantization | Verdict | VRAM | System RAM | Download | Est. write |
|---|---|---|---|---|---|
| Q2_0 | Recommended | 15 GB | 51 GB | ~70 GB | ~130.1 tok/s |
| IQ2_XS | Needs more RAM | 15 GB | 67 GB | ~72 GB | — |
| IQ3_XXS | Needs more RAM | 15 GB | 67 GB | ~74 GB | — |
| IQ3_S | Needs more RAM | 15 GB | 67 GB | ~76 GB | — |
| Coder | Fits | 15 GB | 35 GB | ~68 GB | ~76.1 tok/s |
| Unsloth UD-IQ4_XS | Needs more RAM | 15 GB | 83 GB | ~94 GB | — |
Quick reference: what each VRAM & RAM tier can run (at 32K context)
Every row is computed directly from the same quantization floors and KV-cache rules as the interactive calculator above.
| Hardware tier | GPU VRAM / Unified Memory | System RAM | Recommended quant | Est. write |
|---|---|---|---|---|
| RTX 4060 8GB / RTX 3070 | 8 GB | 32 GB | — | |
| RTX 4070 / RTX 5070 12GB (32 GB RAM) | 12 GB | 32 GB | — | |
| RTX 4070 / RTX 5070 12GB (64 GB RAM) | 12 GB | 64 GB | Q2_0 | ~76 tok/s |
| RTX 5080 / RTX 5070 Ti / RX 9070 XT 16GB | 16 GB | 64 GB | Q2_0 | ~108.5 tok/s |
| RTX 4090 / RTX 3090 / RX 7900 XTX | 24 GB | 64 GB | Q2_0 | ~113.9 tok/s |
| RTX 5090 / 2× RTX 5060 Ti 16GB | 32 GB | 96 GB | Unsloth UD-IQ4_XS | ~97 tok/s |
| Mac Studio M5 Max / Mac mini M5 Pro 64 GB | 64 GB* | 64 GB | Coder | ~36.1 tok/s |
| Ryzen AI Max+ 395 / DGX Spark / Mac Studio 128 GB+ | 128 GB* | 128 GB | Unsloth UD-IQ4_XS | ~43.3 tok/s |
FAQ
- What is the minimum GPU VRAM and RAM to run Strata locally?
- You need at least 12 GB GPU VRAM and 32 GB system RAM for the Coder build, or 12 GB VRAM + 64 GB RAM for the recommended IQ2_XS default build.
- Can an 8 GB GPU like an RTX 4060 run Qwen3.8-Flash-Next?
- No — 8 GB discrete VRAM sits below the 12 GB minimum floor for Qwen3.8-Flash-Next builds. Moving to a 12 GB or 16 GB card (such as RTX 4070, RTX 5070, or RX 9070 XT) unlocks local runs.
- How much extra memory does 128K context require?
- Long 128K context adds roughly 14 GB of KV-cache memory overhead compared to ~1.5 GB at 8K context, so 64 GB–96 GB system RAM (or 128 GB unified memory) is recommended for heavy 128K workloads.
- Is Apple Silicon unified memory faster or slower than an RTX 4090?
- Apple Silicon holds much larger models in one quiet unified pool (64–512 GB), while discrete NVIDIA cards like the RTX 4090 (109.3 tok/s) and RTX 5090 (168.4 tok/s) generate tokens 2–5× faster.