Mac Studio M3 Ultra 96GBBudget pick
96 GB Unified (+48)
C
Adds memory headroom for longer context windows and future model growth.76.1 tok/s decode
Adds memory headroom for longer context windows and future model growth.
~$3,999 MSRP
Can gemma 3 12b it run on Radeon Pro W7900 48GB?
YES — Runs Great
C48Usable
○Estimated from fit model
gemma 3 12b it needs ~14.4 GB VRAM. Radeon Pro W7900 48GB has 48.0 GB. With Q4_K_M quantization, expect ~70 tok/s.
Runtime: llama.cppCapacity: RoomyBandwidth: HighStack: StandardBottleneck: Balanced
Share:
Operating mode
Choose the run profile you care about
Interactive favors responsiveness, while light API and scale-out lean harder on serving readiness. The fit stays the same, but the recommendation lens changes.
Current mode
Balanced
Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.
Select quantization to explore
Q4_K_M (Medium quality) — 14.4 GB, 69.6 tok/s, Runs well
14.4 GB required48.0 GB available
30% VRAM used
Fit status
Runs well
Decode
69.6 tok/s
TTFT
2780 ms
Safe context
398K
Memory
14.4 GB / 48.0 GB
Memory breakdown
Weights7.3 GB
KV Cache1.4 GB
Runtime0.9 GB
Headroom4.8 GB
See how fast it feels
See how fast it feelsgemma 3 12b it on Radeon Pro W7900 48GB
1st promptCold start — includes initialization
>What is local AI inference?
Local AI inference means running an AI language model directly on your own hardware — your laptop, desktop, or server — instead of sending requests to a remote cloud API.
When you run inference locally the model weights are loaded into your GPU or unified memory. Each token you generate requires reading those weights from memory, so memory bandwidth is the main bottleneck for decode speed.
Key benefits of running locally:
- Full privacy: your prompts never leave your machine
- No per-token cost or rate limits
- Works offline once the model is downloaded
- Latency depends only on your hardware
2nd promptWarm — model ready, no init overhead
>How much VRAM do I need?
It depends on the model size and quantization level. A rough rule of thumb:
Model size Q4 (4-bit) Q8 (8-bit) FP16
7B params ~4.3 GB ~7.5 GB ~14 GB
13B params ~7.9 GB ~13.9 GB ~26 GB
70B params ~42.7 GB ~74.9 GB ~140 GB
Most people use 4-bit quantization (Q4_K_M) which gives 90-95% of full quality at a fraction of the memory. A 24 GB GPU can comfortably run most 7B-13B models.
Estimated: 69.6 tok/s decode · 2.8s TTFT (warm) · 174 tok/s prefill
What limits this setup
This setup is broadly balanced for this model.
No major red flags
This recommendation has enough memory headroom and acceptable estimated speed for the selected workload.
Best improvement path
Performance by workload
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | C | Runs well | 69.6 tok/s | 1516 ms | 398K |
| Coding | C | Runs well | 69.6 tok/s | 2780 ms | 398K |
| Agentic Coding | C | Runs well | 69.6 tok/s | 4044 ms | 398K |
| Reasoning | C | Runs well | 69.6 tok/s | 3285 ms | 398K |
| RAG | C | Runs well | 69.6 tok/s | 5055 ms | 398K |
Quantization options
How gemma 3 12b it (12B params) fits at each quantization level on Radeon Pro W7900 48GB (48.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 4.7 GB | Low | C42 |
Q3_K_S | 3 | 5.9 GB | Low | C42 |
NVFP4 | 4 |
Get started
Copy-paste commands to run gemma 3 12b it on your machine.
Run
lms load hf-maziyarpanahi--gemma-3-12b-it-gguf && lms server startUpgrade options