Can I run Llama 3.3 70B on the Apple M4 Pro?
Runs slowlyQ4_K_M runs entirely from 45.2 GB of system RAM: Runs slowly, est. 3.0 tok/s (2.1–3.9, theoretical estimate ±30%).
64 GB unified memory · 273.0 GB/s · FP16 8.1 TFLOPS. Assumes 8k context and an f16 KV cache. Verdicts are shown for each memory size.
Verdict by memory size
The page uses the 64 GB configuration; here is every size.
- 24 GBWon't runQ4_K_M doesn't fit
- 48 GBWon't runTry IQ4_XS: Runs slowlyest. 3.4 tok/s (2.4–4.4, theoretical estimate ±30%)
- 64 GBRuns slowlyest. 3.0 tok/s (2.1–3.9, theoretical estimate ±30%)(used on this page)
Every quant of Llama 3.3 70B on the Apple M4 Pro
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| IQ2_M | 24.12 GB | Runs slowly | est. 7.7 tok/s6.2–9.3calibrated estimate ±20% | 27.4 / 44.8 GB | Unified memory | — |
| Q2_K | 26.38 GB | Runs slowly | est. 7.1 tok/s5.7–8.6calibrated estimate ±20% | 29.6 / 44.8 GB | Unified memory | — |
| IQ4_XS | 37.90 GB | Runs slowly | est. 5.1 tok/s4.1–6.1calibrated estimate ±20% | 41.2 / 44.8 GB | Unified memory | — |
| Q4_K_Mbaseline | 42.52 GB | Runs slowly | est. 3.0 tok/s2.1–3.9theoretical estimate ±30% | 45.2 GB RAM | CPU only | — |
| Q8_0 | 74.98 GB | Won't run | — | Needs about 78.2 GB; 60.0 GB of memory is free (44.8 GB usable by the GPU) | — | — |
Where the memory goes at Q4_K_M
- Weights in system RAM
- 45.2 GB
Context length vs. KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 44.4 GBRuns slowly est. 4.7 tok/s3.8–5.7calibrated estimate ±20% | 43.8 GBRuns slowly est. 4.8 tok/s3.8–5.8calibrated estimate ±20% | 43.4 GBRuns slowly est. 4.8 tok/s3.9–5.8calibrated estimate ±20% |
| 8k | 45.2 GBRuns slowly est. 3.0 tok/s2.1–3.9theoretical estimate ±30% | 44.5 GBRuns slowly est. 4.7 tok/s3.8–5.7calibrated estimate ±20% | 43.9 GBRuns slowly est. 4.8 tok/s3.8–5.8calibrated estimate ±20% |
| 16k | 47.9 GBRuns slowly est. 2.9 tok/s2.0–3.7theoretical estimate ±30% | 45.4 GBRuns slowly est. 3.0 tok/s2.1–3.9theoretical estimate ±30% | 44.7 GBRuns slowly est. 4.7 tok/s3.8–5.7calibrated estimate ±20% |
| 32k | 53.3 GBRuns slowly est. 2.6 tok/s1.8–3.3theoretical estimate ±30% | 48.3 GBRuns slowly est. 2.8 tok/s2.0–3.7theoretical estimate ±30% | 45.6 GBRuns slowly est. 3.0 tok/s2.1–3.9theoretical estimate ±30% |
| 64k | 65.1 GBdoes not fit | 54.0 GBRuns slowly est. 2.5 tok/s1.8–3.3theoretical estimate ±30% | 48.6 GBRuns slowly est. 2.8 tok/s2.0–3.6theoretical estimate ±30% |
| 128k | 87.2 GBdoes not fit | 67.2 GBdoes not fit | 54.8 GBRuns slowly est. 2.5 tok/s1.7–3.2theoretical estimate ±30% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At Q4_K_M the verdict stays Runs slowly up to 32k context with an f16 KV cache, and up to 64k with q8_0.
Estimated speed
- Generation
- est. 3.0 tok/s (2.1–3.9, theoretical estimate ±30%)
- Prompt processing
- Prompt-processing estimates only apply when the whole model runs on the GPU.
Measured on this exact combination
No public measurement for this exact combination yet; the numbers above are estimates.
If this is not enough
A smaller model that runs well on the Apple M4 Pro
Runs greatK2-Horizon MoVA 36B-A4B: Runs great, est. 36.8 tok/s (25.8–47.8, theoretical estimate ±30%)
Llama 3.3 70B on the Apple M4 Max (40-core GPU)
Runs wellQ4_K_M: Runs well, est. 9.2 tok/s (7.3–11.0, calibrated estimate ±20%)
How to run it
Commands for the Q4_K_M file. Both tools download from Hugging Face on first run.
File: Llama-3.3-70B-Instruct-Q4_K_M.gguf (42.52 GB) from bartowski/Llama-3.3-70B-Instruct-GGUF.
llama-server -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M -c 8192 -ngl 0 -fa on- -ngl 0 keeps everything on the CPU.
- -fa on enables flash attention (needed for KV cache quantization).
- --cache-type-k q8_0 --cache-type-v q8_0 shrinks the KV cache to 1.44 GB at 8k context.
Terminal:
OLLAMA_CONTEXT_LENGTH=8192 ollama serve
ollama run hf.co/bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M- Ollama defaults to 4k context on GPUs under 24 GB. Set OLLAMA_CONTEXT_LENGTH when starting the server, or /set parameter num_ctx inside the chat.
/set parameter num_ctx 8192
Related pages
Guide: Running local LLMs on a Mac (M4, M5): how much unified memory?
Other models on the Apple M4 Pro
Llama 3.3 70B on other hardware
Frequently asked questions
Can I run Llama 3.3 70B on the Apple M4 Pro?
Q4_K_M runs entirely from 45.2 GB of system RAM: Runs slowly, est. 3.0 tok/s (2.1–3.9, theoretical estimate ±30%). Q4_K_M needs 45.2 GB at 8k context; the GPU can use 44.8 GB of the unified memory.
How much context can Llama 3.3 70B use on the Apple M4 Pro?
At Q4_K_M the verdict stays Runs slowly up to 32k context with an f16 KV cache (10.7 GB of KV) and up to 64k with a q8_0 KV cache (11.5 GB). The model supports up to 128k.
Which quant should I use, and how fast is it?
Q4_K_M (42.52 GB) is the recommended balance: Runs slowly, est. 3.0 tok/s (2.1–3.9, theoretical estimate ±30%). It is also the largest tracked file that gets this verdict.
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.