Cerebras-GPT 13B needs ~19.9 GB VRAM. Mac mini M2 24GB has 17.3 GB. With Q3_K_S quantization, expect ~8 tok/s.
Operating mode
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
5.6 GB over capacity — needs offload or smaller quantization
Fit status
Too heavy
Decode
4.8 tok/s
TTFT
40530 ms
Safe context
7K
Memory
22.9 GB / 17.3 GB
Offload
20%
It fits through host-memory offload, and offload is the main reason performance drops.
CPU or host-memory offload is active
About 10% of the working set spills out of accelerator memory, which usually hurts latency and sustained decode throughput.
Very little memory headroom
You can run the model, but there is not much room left for longer context, bigger batches, extra apps, or future model updates.
Shared-memory contention still exists
The OS, browser, and inference runtime all compete for the same physical memory pool, so real-world headroom is less forgiving than raw capacity suggests.
Remove offload with more accelerator memory
Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.
Buy headroom, not only minimum fit
A slightly larger memory tier gives you safer context growth and makes the recommendation more future-proof.
Increase host RAM if you keep offloading
This setup may need roughly 0.8 GB of extra host RAM just for the offloaded portion, before OS and other tools.
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | B | Runs with offload (needs ~0.4 GB host RAM) | 6.5 tok/s | 16186 ms | 7K |
| Coding | F | Too heavy | 4.8 tok/s | 40530 ms | 7K |
| Agentic Coding | F | Too heavy | 3.2 tok/s | 87917 ms | 7K |
| Reasoning | F | Too heavy | 4.8 tok/s | 47899 ms | 7K |
| RAG | F | Too heavy | 3.2 tok/s | 109896 ms | 7K |
Inference speed
Estimated decode speed (tokens/sec) for Cerebras-GPT 13B at Q5_K_M across popular GPUs and Apple Silicon, using the fastest local runtime per device. Fastest is RTX 5090 32GB at ~131 tok/s. Speed is memory-bandwidth bound, so cards that fit the whole model in VRAM run far faster than ones that offload to system RAM.
| GPU / Mac | Memory | Quant | Speed (tok/s) | Fits? |
|---|---|---|---|---|
 RTX 5090 32GB | 32 GB | Q5_K_M | 130.8 | Fits |
| RTX 4090 24GB | 24 GB | Q5_K_M | 83.5 | Tight |
RX 7900 XTX 24GB | 24 GB | Q5_K_M | 75.3 | Tight |
| RTX 3090 24GB | 24 GB | Q5_K_M | 71.4 | Tight |
Mac Studio M3 Ultra 256GB | 256 GB | Q5_K_M | 60.7 | Fits |
Mac Studio M2 Ultra 128GB | 128 GB | Q5_K_M | 50.6 | Fits |
Mac Studio M1 Ultra 128GB | 128 GB | Q5_K_M | 47.9 | Fits |
MacBook Pro M4 Max 128GB | 128 GB | Q5_K_M | 33.0 | Fits |
MacBook Pro M4 Max 64GB | 64 GB | Q5_K_M | 33.0 | Fits |
MacBook Pro M3 Max 64GB | 64 GB | Q5_K_M | 26.2 | Fits |
| RTX 4080 Super 16GB | 16 GB | Q5_K_M | 25.7 | Too big |
MacBook Pro M1 Max 64GB | 64 GB | Q5_K_M | 24.0 | Fits |
MacBook Pro M4 Pro 48GB | 48 GB | Q5_K_M | 20.2 | Fits |
| RTX 4070 12GB | 12 GB | Q5_K_M | 9.0 | Too big |
| RTX 3060 12GB | 12 GB | Q5_K_M | 5.7 | Too big |
| RTX 4060 8GB | 8 GB | Q5_K_M | 3.2 | Too big |
Estimates for single-stream decoding at Q5_K_M; real tokens/sec varies with prompt length, context, batch size, and runtime build. Prompt processing (prefill) is faster than the decode figures shown here.
How Cerebras-GPT 13B (13B params) fits at each quantization level on Mac mini M2 24GB (17.3 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 5.1 GB | Low | B64 |
Q3_K_S | 3 | 6.4 GB | Low | B65 |
NVFP4 | 4 | 7.3 GB | Medium | B66 |
Q4_K_M | 4 | 7.9 GB | Medium | B67 |
Q5_K_M | 5 | 9.4 GB | High | B67 |
Q6_KBest for your GPU | 6 | 10.7 GB | High | B67 |
Q8_0 | 8 | 13.9 GB | Very High | F0 |
F16 | 16 | 26.7 GB | Maximum | F0 |
Copy-paste commands to run Cerebras-GPT 13B on your machine.
Run
docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
--hf-repo "cerebras/Cerebras-GPT-13B" \
--hf-file "Cerebras-GPT-13B-Q5_K_M.gguf" \
-c 4096 -ngl 99升级选项
Makes the model fit on the accelerator instead of staying completely out of reach.
Removes host-memory offload, which is usually the single biggest latency and throughput win.
~$799 MSRP
Makes the model fit on the accelerator instead of staying completely out of reach.
Removes host-memory offload, which is usually the single biggest latency and throughput win.
~$1,099 MSRP
Makes the model fit on the accelerator instead of staying completely out of reach.
Removes host-memory offload, which is usually the single biggest latency and throughput win.
~$1,099 MSRP
RTX 5090 32GB最大飞跃Makes the model fit on the accelerator instead of staying completely out of reach.
Removes host-memory offload, which is usually the single biggest latency and throughput win.
~$1,999 MSRP
Yes, Mac mini M2 24GB can run Cerebras-GPT 13B at Q3_K_S quantization (Very compromised (needs ~0.8 GB host RAM)). The recommended Q5_K_M requires 22.9 GB which exceeds available memory, but at Q3_K_S it needs only 19.9 GB. Expected decode speed: 7.6 tok/s.
Cerebras-GPT 13B (13B parameters) requires approximately 22.9 GB at Q5_K_M quantization. On Mac mini M2 24GB, it fits at Q3_K_S using 19.9 GB.
The recommended quantization is Q5_K_M, but on Mac mini M2 24GB the best fitting quantization is Q3_K_S, which uses 19.9 GB.
On Mac mini M2 24GB, Cerebras-GPT 13B achieves approximately 7.6 tokens per second decode speed with a time-to-first-token of 25467ms using Q3_K_S quantization.
For coding workloads, Cerebras-GPT 13B on Mac mini M2 24GB receives a F grade with 4.8 tok/s and 7K context.
On Mac mini M2 24GB, Cerebras-GPT 13B can safely use up to 12K tokens of context at Q3_K_S quantization. The model's official context limit is 131K, but available memory constrains the safe maximum.
Remove offload with more accelerator memory. Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.
Not always. Mac mini M2 24GB can often fit larger models thanks to unified memory, but a discrete GPU with dedicated high-bandwidth VRAM may still decode faster once the model fits. For this combination, the important distinction is capacity versus sustained throughput.
Paste this snippet into any page to show a live fit card.
<iframe src="https://willitrunai.com/embed/cerebras-gpt-13b-on-m2-24gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>
Preview: