What LLM can I run with 8 GB of VRAM?
At the default reserve, one 8 GB GPU clears an 8B INT4, 3B INT8, or 3B FP16 weight floor. The exact checkpoint and context can still exceed that budget.
Instrument 16 / hardware-first model fit
What LLM can I run on my GPU?
Start with the VRAM you already own. Find the largest standard model-size tier that clears a conservative weight floor, then add the exact architecture and context before downloading the checkpoint.
Find the model-size floor →First answer / capacity floor
Turn available VRAM into a model-size ceiling.
The calculation stays in this browser. It answers the fully GPU-resident weight question first and deliberately excludes KV cache, artifact metadata, and speed.
Read the answer correctly
Tier, not recommendation.
A parameter class clearing the floor does not name the best model for your task.
Weights, not full runtime.
KV cache, activations, workspace, metadata, and fragmentation still consume memory.
Capacity, not speed.
Bandwidth, kernels, offload, and interconnect decide delivered performance.
VRAM-first planning table
What model-size floor fits 8, 12, 16, or 24 GB?
Each cell is the largest listed tier below the arithmetic ceiling with 20% runtime headroom, 90% usable memory, one GPU, and no KV cache. The small number is the ceiling before rounding down to a standard tier.
Swipe the matrix horizontally →
| VRAM profile | INT4 | INT8 | FP16 / BF16 |
|---|---|---|---|
| 8 GB | 8B tier12.9B ceiling | 3B tier6.4B ceiling | 3B tier3.2B ceiling |
| 12 GB | 14B tier19.3B ceiling | 8B tier9.7B ceiling | 3B tier4.8B ceiling |
| 16 GB | 20B tier25.8B ceiling | 8B tier12.9B ceiling | 3B tier6.4B ceiling |
| 24 GB | 32B tier38.7B ceiling | 14B tier19.3B ceiling | 8B tier9.7B ceiling |
| 32 GB | 32B tier51.5B ceiling | 20B tier25.8B ceiling | 8B tier12.9B ceiling |
| 48 GB | 70B tier77.3B ceiling | 32B tier38.7B ceiling | 14B tier19.3B ceiling |
| 80 GB | 120B tier128.8B ceiling | 32B tier64.4B ceiling | 32B tier32.2B ceiling |
| 141 GB | 120B+ tier227.1B ceiling | 70B tier113.5B ceiling | 32B tier56.8B ceiling |
The table uses ideal bit-width arithmetic. A real quantized artifact can be larger because tensors, metadata, embeddings, or selected layers may use different precision.
Transparent method
The result is a first gate, not a runtime promise.
Usable capacity is GPU memory × GPU count × the usable-memory percentage you select.
Weight capacity divides usable capacity by one plus the selected runtime headroom. The remaining GiB are converted into parameters at the chosen bit width.
Largest listed tier rounds down to one of the common planning sizes on this page: 3B, 7B, 8B, 13B, 14B, 20B, 27B, 32B, 70B, or 120B.
Exact deployment memory needs checkpoint size, layer count, KV heads, head dimension, context length, concurrency, and cache precision. The result button carries your selected tier and hardware into that worksheet.
Multi-GPU capacity is summed for the floor only. The model and serving runtime must support a workable tensor, pipeline, or expert-parallel layout.
Named models vary inside a tier. Use the primary-source worksheets for Qwen3 8B, 14B, and 32B, gpt-oss 20B and 120B, or the 70B worksheet.
Know the model instead? Check a model against your GPU →Capacity references
Verify the exact card before using a preset.
The finder is vendor-neutral. These first-party hardware pages are capacity anchors for common profiles; board variants and software support can differ.
| Profile | First-party reference | Usable at 90% |
|---|---|---|
| 12 GB | NVIDIA GeForce RTX 3060 | 10.8 GiB |
| 24 GB | NVIDIA GeForce RTX 4090 | 21.6 GiB |
| 32 GB | NVIDIA GeForce RTX 5090 | 28.8 GiB |
| 48 GB | NVIDIA RTX 6000 Ada | 43.2 GiB |
| 80 GB | NVIDIA A100 80GB / H100 SXM | 72 GiB |
| 141 GB | NVIDIA H200 | 126.9 GiB |
GPU model-fit FAQ
The questions behind “will it run?”
The fast answer is a capacity floor. The deployment answer comes from the exact artifact, context, runtime, and performance target.
What LLM can I run with 12 GB of VRAM?
At the default reserve, one 12 GB GPU clears a 14B INT4, 8B INT8, or 3B FP16 weight floor. Use the exact architecture calculator before downloading a model.
What LLM can I run with 24 GB of VRAM?
At the default reserve, one 24 GB GPU clears a 32B INT4, 14B INT8, or 8B FP16 weight floor. This does not guarantee speed, kernel support, or a desired context window.
Why can a model fail even when its weights fit in VRAM?
Serving also needs KV cache, activations, workspace, metadata, and fragmentation reserve. Context, batching, quantization format, supported kernels, and topology can change the boundary.
Evidence before the download
Get the next infrastructure teardown.
ResearchAudio traces architecture, measured constraints, primary sources, and the gap between “fits” and “works.”