Model → hardware
Gemma 3 12B
Plan for 12 GB of dedicated GPU memory at Q4_K_M and a 4,000-token context. 16 GB gives more headroom under the same settings.
Text inference only. Vision needs a matching projector and additional memory; it is outside these estimates.
Text inference with Gemma 3. Memory fit is an estimate, not a speed or quality guarantee.
Hardware starting points
Planning estimatesQ4_K_M · 32 GB system RAM · one active request · full GPU residency
| Dedicated GPU memory | At 4,000 tokens | At 16,000 tokens | Example from these guides |
|---|---|---|---|
| 12 GB | Smallest listed tier that fits9.7 GB + 0.36 GB reserve | Fits the estimate10.6 GB + reserve | RTX 3060 12 GB |
| 16 GB | More headroom9.7 GB + 0.48 GB reserve | Fits the estimate10.6 GB + reserve | RTX 4060 Ti 16 GB |
The smallest tier fits the calculation; it is not a verified minimum. Cache precision and runtime buffers can change the fit. “More headroom” means at least 2 GB remains after the estimate and reserve.
The download is a GGUF conversion; Q5 in the table is an estimate, not a listed file in that repository. Use the publisher’s Instruct / chat template in a current GGUF-capable runtime. The 12 GB tier can become tight when display use, buffers or context grow. CPU offloading needs adequate system RAM and changes performance; it is not included in these full-GPU estimates.
Compare quantisation and context
| File profile | Weight input¹ | GPU estimate at 4K | GPU estimate at 16K |
|---|---|---|---|
| Q4_K_M | 8.50 GB | 9.7 GB | 10.6 GB |
| Q5_K_M | 10.40 GB | 11.6 GB | 12.5 GB |
¹ Catalogue planning weights. These can differ from an exact downloaded file; runtime memory also includes cache and buffers.
Keep the exact file, runtime version, cache type and context limit consistent when comparing. Q5 uses more memory; compare its answers on your own tasks before accepting the trade-off.
Setup and common questions
Does this estimate include images?
No. Gemma-3-12b-it accepts text and images, but this page estimates text inference only. In llama.cpp, vision also needs a matching multimodal projector and additional buffers. Measure that configuration separately; a text fit is not a vision fit.
Which quantisation matches the table?
ggml-org provides a Q4_K_M GGUF conversion, about 7.30 GB before cache and buffers. The Finder uses an 8.50 GB weight input. Its Q5 table is a planning profile; the linked repository has no Q5 file. Google’s QAT Q4_0 edition is a different profile.
What should I check before loading it?
Use a current runtime with Gemma 3 support and the correct chat template. The 128K model window does not guarantee a hardware fit. Google’s Gemma Terms apply to the checkpoint and its distribution.
How are these recommendations chosen?
We compare dedicated GPU memory tiers using the main Finder’s memory calculation at the stated quantisation and context. These tiers describe memory fit, not speed or output quality.
GPU memory includes weights, context cache and a 0.9 GB runtime allowance, plus a 3% reserve (0.15–1.5 GB). Cache format and runtime buffers can change the fit. See the calculation assumptions.
Sources and related guides
Specification sources checked 3 October 2026.
These comparisons are planning estimates. No reviewed or independently reproduced measurements are available yet. About test evidence.