Mac Studio M3 Ultra 256GBBudget pick
256 GB Unified (+128)
D
Adds memory headroom for longer context windows and future model growth.4.2 tok/s decode
Adds memory headroom for longer context windows and future model growth.
~$6,999 MSRP
Can embeddinggemma 300M run on AMD Instinct MI250X 128GB?
YES — Runs Great
D35Poor
○Estimated from fit model
embeddinggemma 300M needs ~14.0 GB VRAM. AMD Instinct MI250X 128GB has 128.0 GB. With Q6_K quantization, expect ~4 tok/s.
Runtime: llama.cppCapacity: RoomyBandwidth: HighStack: StandardBottleneck: Memory bandwidth
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
Q6_K (High quality) — 14.0 GB, 4.2 tok/s, Runs well
14.0 GB required128.0 GB available
11% VRAM used
Fit status
Runs well
Decode
4.2 tok/s
TTFT
46095 ms
Safe context
18.2M
Memory
14.0 GB / 128.0 GB
Memory breakdown
Weights0.2 GB
KV Cache0.1 GB
Runtime0.9 GB
Headroom12.8 GB
See how fast it feels
See how fast it feelsembeddinggemma 300M on AMD Instinct MI250X 128GB
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: 4.2 tok/s decode · 46.1s TTFT (warm) · 11 tok/s prefill
What limits this setup
This model fits, but memory bandwidth is the part holding decode speed back.
Throughput will feel slow
Estimated decode speed is only 4.2 tok/s, so this is more of a technical fit than a comfortable daily-driver setup.
Best improvement path
Prioritize bandwidth, not only capacity
If this workload feels slow, the next useful step is often a GPU tier with materially faster memory bandwidth rather than only a small bump in capacity.
Performance by workload
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | D | Runs well | 4.2 tok/s | 25143 ms | 9.1M |
| Coding | D | Runs well | 4.2 tok/s | 46095 ms | 18.2M |
| Agentic Coding | D | Runs well | 4.2 tok/s | 67048 ms | 36.5M |
| Reasoning | D | Runs well | 4.2 tok/s | 54476 ms | 18.2M |
| RAG | D | Runs well | 4.2 tok/s | 83810 ms | 36.5M |
Quantization options
How embeddinggemma 300M (0.30000001192092896B params) fits at each quantization level on AMD Instinct MI250X 128GB (128.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 0.1 GB | Low | D38 |
Q3_K_S | 3 | 0.1 GB | Low | D38 |
NVFP4 | 4 |
Get started
Copy-paste commands to run embeddinggemma 300M on your machine.
Run
docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
--hf-repo "ggml-org/embeddinggemma-300M-GGUF" \
--hf-file "embeddinggemma-300M-GGUF-Q6_K.gguf" \
-c 4096 -ngl 99Upgrade options