MacBook Pro M4 Max 40c · 128 GB
Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
- llama-2-7b · macbook-pro-m4-max-128g
llamafile (Metal) · Q4_0 · LocalScore generation
Recorded 2026-10-12 · historical source record · repeat count unknownSourced
Third-party measurement used to calibrate η (brief §4.3): implies η ≈ 0.43. Run source - llama-2-7b · macbook-pro-m4-max-128g
llama.cpp (Metal) · Q4_0 · tg128
Recorded 2026-10-12 · historical source record · repeat count unknownSourced
Llama 2 7B Q4_0, llama-bench tg128, M4 Max 40-core GPU (546 GB/s), pinned commit 8e672ef (Nov 2023) for comparability; newer builds are faster. Run source
Evidence labels · Freshness policy · Calculation methodology
A laptop with twice Strix Halo's bandwidth. MLX and Metal llama.cpp are both mature.
Decode-speed ceiling for a model is 546 GB/s ÷ GB read per token. For a 16.6 GB file that is 33 tok/s; real runtimes reach roughly 45% of it. Why bandwidth is the limit.
Specifications
Memory | 128 GBSourcedConfigurable to 128 GB on the 16-core-CPU/40-core-GPU M4 Max, Apple tech specs, checked 2026-10-12. Source |
|---|---|
Bandwidth | 546 GB/sSourcedVendor spec page, checked 2026-10-12. Source |
FP16 dense compute | 34 TFLOPSSampleSample FP16 dense compute; pending vendor-sheet citation. |
Load power at wall | 140 WSampleSample wall power under load. |
Price, USD | $4,699SampleSample price. |
Price, CAD | CA$6,706SampleUSD price × 1.4271 (Bank of Canada USD/CAD, 2026-10-09); no Canadian retail listing captured. Converted from a sample USD value. Source |
Models that run
Default quant, single stream, 8k context. Fastest first; estimates are ±25% until measured.
Model | Quant | Memory | Verdict | Decode, tok/s | Max context |
|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 | 12.9 GB | Fits | 107.3Estimate | 128k |
| Qwen3 30B-A3B | Q4_K_M | 19.9 GB | Fits | 87.4Estimate | 32k |
| gpt-oss-120b | MXFP4 | 64.3 GB | Fits | 79.9Estimate | 128k |
| Llama 3.1 8B Instruct | Q4_K_M | 6.6 GB | Fits | 41.0Estimate | 128k |
| Qwen3 8B | Q4_K_M | 6.8 GB | Fits | 39.4Estimate | 32k |
| Qwen3 14B | Q4_K_M | 10.9 GB | Fits | 23.8Estimate | 32k |
| Mistral Small 3.2 24B | Q4_K_M | 16.2 GB | Fits | 15.7Estimate | 128k |
| Gemma 3 27B IT | Q4_K_M | 18.2 GB | Fits | 13.9Estimate | 128k |
| Qwen3 32B | Q4_K_M | 22.5 GB | Fits | 11.2Estimate | 32k |
| Llama 3.3 70B Instruct | Q4_K_M | 45.8 GB | Fits | 5.4Estimate | 128k |
Published reference runs
Third-party llama.cpp measurements on this device, with conditions and source. Llama 2 7B Q4_0 runs are the community calibration test; they are not catalogue models.
Model | Quant | Date | Runtime · test | Decode, tok/s | Conditions |
|---|---|---|---|---|---|
| llama-2-7b | Q4_0 | 2026-10-12 | llamafile (Metal) · LocalScore generation | 62.5Sourced | Third-party measurement used to calibrate η (brief §4.3): implies η ≈ 0.43. Source |
| llama-2-7b | Q4_0 | 2026-10-12 | llama.cpp (Metal) · tg128 | 83.1Sourced | Llama 2 7B Q4_0, llama-bench tg128, M4 Max 40-core GPU (546 GB/s), pinned commit 8e672ef (Nov 2023) for comparability; newer builds are faster. Source |