ResearchAudio

Hardware guide / 12 GB VRAM

What LLM can an RTX 3060 12GB run?

A 12 GB card has a useful 8B INT4 lane, but context memory matters. Qwen3 8B at 32K clears this planning model; Qwen3 14B at the same context does not.

Open the 12 GB model finder →

Fast answer

The generic weight-only floor

INT4
14B listed tier · 19.33B arithmetic ceiling

INT8
8B listed tier · 9.66B arithmetic ceiling

FP16 / BF16
3B listed tier · 4.83B arithmetic ceiling

Default reserve
12 × 0.90 ÷ 1.20 = 9 GiB for weights

Editable evidence

Where the 12 GB boundary actually lands

Every card opens the calculator with the architecture, context, reserve, and headroom already filled in.

Qwen3 8B · INT4 · 32K9.98 GiBFits with about 0.82 GiB margin → Qwen3 8B · INT8 · 32K14.56 GiBExceeds the 10.8 GiB budget → Qwen3 14B · INT4 · 32K14.27 GiBWeights fit; this full profile does not → gpt-oss-20b · MXFP4 · 4K14.29 GiBOfficial checkpoint exceeds the budget → Generic 20B · INT4 · weights only11.18 GiBEven the weight-only plan narrowly misses → RTX 5080 comparison14.4 GiBSee what the 16 GB tier changes →

Decision boundary

12 GB works best when the context is explicit.

The generic tier finder sizes weights only. The named-model profiles add the KV cache and an explicit runtime reserve.

ProfilePlanning target12 GB result
Qwen3 8B INT4 · 32K9.98 GiBFits with 0.82 GiB margin
Qwen3 8B INT8 · 32K14.56 GiBDoes not fit
Qwen3 14B INT4 · 32K14.27 GiBDoes not fit fully on GPU
gpt-oss-20b · 4K14.29 GiBDoes not fit under this reserve

Capacity is only gate one

Weights are not the deployment.
Qwen3 14B adds 5 GiB of BF16 KV cache at 32K before runtime headroom.

Context consumes VRAM.
Shorter context or KV-cache quantization can move the boundary.

Fit is not throughput.
Bandwidth, kernels, offload, and batching decide delivered speed.

Primary sources

Inputs you can inspect.

NVIDIA documents both 12 GB and 8 GB RTX 3060 variants. This guide is specifically for the 12 GB GDDR6 card. Model profiles reuse the primary-source inputs documented in the Qwen3 and gpt-oss worksheets.

Official NVIDIA RTX 3060 family specifications → Open the Qwen3 GPU worksheet → Open the gpt-oss hardware worksheet →

What LLM can an RTX 3060 12GB run?

At the default reserve, the generic one-card floor reaches a 14B INT4, 8B INT8, or 3B FP16 tier. Named models still need checkpoint, context, cache, and runtime inputs.

Can an RTX 3060 12GB run Qwen3 8B?

The 32K INT4 planning profile reaches 9.98 GiB and clears the 10.8 GiB budget by about 0.82 GiB. Validate the exact artifact and runtime.

Can an RTX 3060 12GB run Qwen3 14B?

Its raw four-bit weights can fit, but the fully GPU-resident 32K profile reaches 14.27 GiB. Shorter context, KV-cache quantization, or offload can move the boundary.

Can an RTX 3060 12GB run gpt-oss-20b?

Not under this conservative full-GPU plan. The official 12.8 GiB checkpoint already exceeds the 10.8 GiB usable allowance.

Hardware claims, checked

Get the next local-AI deployment teardown.

ResearchAudio traces exact checkpoints, architecture, context, runtime constraints, and the gap between “fits” and “works.”