Apple M4 Pro for local LLMs
Quick answer
How fast is the Apple M4 Pro for local LLMs?
The fastest plain-decode result we have is 127.9 tok/s on Flint 40B-A2.5B (int4 MLX (RecursiveSI/flint-int4), Kapitsa (PR #286)) in MacBook Pro M4 Pro 48 GB: Flint 40B-A2.5B 4-bit at 127.9 tok/s (Kapitsa), 126.6 (rec), 79.6 (llama.cpp). Speed depends heavily on the model, quant and engine, so compare like with like in the table. RigBase lists 7 measured results from 3 builds.
Last updated:
7 measured results from 3 builds, each linked to its source.
| Memory | unified memory |
|---|---|
| Memory bandwidth | not listed |
| Typical EU price | not tracked yet |
Can it run? Flint 40B-A2.5B on Apple M4 Pro Qwen3.8 27B on Apple M4 Pro Qwen3.8 Flash-Next on Apple M4 Pro
All builds with the Apple M4 Pro
| Build | GPU(s) | Memory GB (total) | Results |
|---|---|---|---|
| MacBook Pro M4 Pro 48 GB: Flint 40B-A2.5B 4-bit at 127.9 tok/s (Kapitsa), 126.6 (rec), 79.6 (llama.cpp) | Apple M4 Pro (20-core GPU, unified) | 48 | 3 |
| MacBook Pro M4 Pro 24 GB: Qwen3.8 27B GSQ IQ3_S (12.1 GB) on llama.cpp Metal at 11.4 tok/s, MTP gives no gain (Kingy) | Apple M4 Pro (unified, MacBook Pro) | 24 | 2 |
| Mac mini M4 Pro 24 GB: Qwen3.8 Flash-Next (125B-A6B) IQ3_XXS at 10.7 tok/s, 9.4 after a 6K prompt (MoEShift, SSD experts) | Apple M4 Pro (unified) | 24 | 2 |
All measured results
(MTP, DFlash, DSpark and similar draft-and-verify methods give much higher tok/s on predictable text and are not comparable 1:1 with plain decode)
| Model | Build | Ran on | Quant | Engine | tok/s | Details |
|---|---|---|---|---|---|---|
| Flint 40B-A2.5B | MacBook Pro M4 Pro 48 GB: Flint 40B-A2.5B 4-bit at 127.9 tok/s (Kapitsa), 126.6 (rec), 79.6 (llama.cpp) | Apple M4 Pro | int4 MLX (RecursiveSI/flint-int4) | Kapitsa (PR #286) | 127.9 | |
| Flint 40B-A2.5B | MacBook Pro M4 Pro 48 GB: Flint 40B-A2.5B 4-bit at 127.9 tok/s (Kapitsa), 126.6 (rec), 79.6 (llama.cpp) | Apple M4 Pro | int4 MLX (same files) | rec 0.1.0 (Recursive) | 126.56 | |
| Flint 40B-A2.5B | MacBook Pro M4 Pro 48 GB: Flint 40B-A2.5B 4-bit at 127.9 tok/s (Kapitsa), 126.6 (rec), 79.6 (llama.cpp) | Apple M4 Pro | Q4_1 GGUF (RecursiveSI/flint-int4-gguf) | llama.cpp b11804, all layers on GPU, -fa on | 79.57 | |
| Qwen3.8 27B | MacBook Pro M4 Pro 24 GB: Qwen3.8 27B GSQ IQ3_S (12.1 GB) on llama.cpp Metal at 11.4 tok/s, MTP gives no gain (Kingy) | Apple M4 Pro | GSQ-RCO IQ3_S + MTP head (ISTA-DASLab GGUF, 12.12 GB) | llama.cpp 0.6.0 (Metal) | 11.4 | context: 32K ctx, one request at a time📏 Kingy test 8 Nov 2026: llama-server, 32K ctx, prompt cache off, temperature 0, seed 42, up to 1,024 reply tokens (all spent thinking); average of 4 prompts (code, explainer, HTML report, question on an 8,217-token document); MTP off; llama-bench tg128 12.3 (10.9 with 8K in context); about 82-87 s before the first token on the 8K document |
| Qwen3.8 27B | MacBook Pro M4 Pro 24 GB: Qwen3.8 27B GSQ IQ3_S (12.1 GB) on llama.cpp Metal at 11.4 tok/s, MTP gives no gain (Kingy) | Apple M4 Pro | GSQ-RCO IQ3_S + MTP head (ISTA-DASLab GGUF, 12.12 GB) | llama.cpp 0.6.0 (Metal) | 10.9 | ⚡ spec-decode (MTP)context: 32K ctx, one request at a time📏 Kingy test 8 Nov 2026: llama-server, 32K ctx, prompt cache off, temperature 0, seed 42, up to 1,024 reply tokens (all spent thinking); average of 4 prompts (code, explainer, HTML report, question on an 8,217-token document); --spec-type draft-mtp --spec-draft-n-max 2; 70% of drafts kept but no speed gain on this Mac |
| Qwen3.8 Flash-Next | Mac mini M4 Pro 24 GB: Qwen3.8 Flash-Next (125B-A6B) IQ3_XXS at 10.7 tok/s, 9.4 after a 6K prompt (MoEShift, SSD experts) | Apple M4 Pro | ISTA-DASLab GSQ-RCO IQ3_XXS (~76 GB, 2 shards), experts streamed from the internal SSD | MoEShift 0.1.0-alpha.5 (llama.cpp-based, Metal) | 10.74 | context: 72-token prompt, 512 generated (64K ctx configured)📏 moeshift-community-v1 protocol: median of 3 measured runs after 1 warm-up, 512 output tokens, prompt cache off (cache_n 0); short-prompt generation runs 10.74-10.78partly in system RAM |
| Qwen3.8 Flash-Next | Mac mini M4 Pro 24 GB: Qwen3.8 Flash-Next (125B-A6B) IQ3_XXS at 10.7 tok/s, 9.4 after a 6K prompt (MoEShift, SSD experts) | Apple M4 Pro | ISTA-DASLab GSQ-RCO IQ3_XXS (~76 GB, 2 shards), experts streamed from the internal SSD | MoEShift 0.1.0-alpha.5 (llama.cpp-based, Metal) | 9.38 | context: 6,060-token prompt, 512 generated📏 moeshift-community-v1 protocol: median of 3 measured runs after 1 warm-up, 512 output tokens, prompt cache off (cache_n 0); long-prompt runs 9.37-9.38, prefill 197.2-199.9partly in system RAM |
All Apple M4 Pro builds: see also all GPUs · best LLMs by memory size · download the dataset.
FAQ
How fast is the Apple M4 Pro for local LLMs?
The fastest plain-decode result we have is 127.9 tok/s on Flint 40B-A2.5B (int4 MLX (RecursiveSI/flint-int4), Kapitsa (PR #286)) in MacBook Pro M4 Pro 48 GB: Flint 40B-A2.5B 4-bit at 127.9 tok/s (Kapitsa), 126.6 (rec), 79.6 (llama.cpp). Speed depends heavily on the model, quant and engine, so compare like with like in the table. RigBase lists 7 measured results from 3 builds.
How much memory does the Apple M4 Pro have?
unified memory.
Which models have been measured on the Apple M4 Pro?
Flint 40B-A2.5B, Qwen3.8 27B, Qwen3.8 Flash-Next.
Where do these numbers come from?
Every row links to a build page with its original public source (benchmark logs, READMEs, forum posts). RigBase does not re-run the benchmarks. Results marked ⚡ use speculative decoding and are not comparable 1:1 with plain decode.
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.