selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

nvidia logo

Can the Apple M4 (16 GB) run NLD-Uni?

NLD-Uni·18.3B·Apple M4 (16 GB)·16 GB unified memory

Saved in this browser, used on every model page.
Just about

Just about. At Q3_K_M it needs about 10 GB of the 11 GB available, so keep the context short, at roughly 7.4 tokens/s.

best quant
Q3_K_M
needs
10 GB
usable
11 GB
est. speed
~7.4 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the NLD-Uni page →

Every quantisation of NLD-Uni on the Apple M4 (16 GB)

QuantQualityMemoryFitSpeed
Q2_K Heavy loss7.4 GBfits~11 tok/s
Q3_K_M pickNoticeable loss10 GBtight~7.4 tok/s
Q4_K_M Small loss12 GBtoo big—
Q5_K_M Near-lossless14 GBtoo big—
Q6_K Near-lossless16 GBtoo big—
Q8_0 Lossless21 GBtoo big—
FP16 / BF16 Lossless37 GBtoo big—

Speed is a bandwidth-based estimate for generating text (120 GB/s); prompt processing and long contexts are slower.

Measured speed on the Apple M4 (16 GB)

No community reports for this pair yet.

Report your own speed →

Omni models that fit the Apple M4 (16 GB) comfortably

All that fit →
ModelParamsBest quantMemorySpeedLicence
ByteDance-Seed logoBAGEL-7B-MoTByteDance-Seed
14.7BQ4_K_M≈9.6 GBFits in memory~7.4 tok/sCommercial OK
10.7BQ8_0≈8.9 GBFits in memory~5.8 tok/sCommercial OK
9.4BQ8_0≈10 GBFits in memory~6.6 tok/sCommercial OK
8BQ8_0≈9.2 GBFits in memory~7.8 tok/sCommercial OK
9.1BQ8_0≈10 GBFits in memory~6.9 tok/sCommercial OK
8.3BQ8_0≈10 GBFits in memory~7.5 tok/sCheck licence

Other GPUs that run NLD-Uni

Frequently asked questions

Can the Apple M4 (16 GB) run NLD-Uni?
Just about. At Q3_K_M it needs about 10 GB of the 11 GB available, so keep the context short, at roughly 7.4 tokens/s.
How fast does NLD-Uni run on the Apple M4 (16 GB)?
An estimate from memory bandwidth gives about 7.4 tokens/s at Q3_K_M.
Which quantisation of NLD-Uni should I use on the Apple M4 (16 GB)?
Q3_K_M: the largest one that fits. Expected quality: noticeable loss compared with the full-precision weights.

Memory and speed are estimates from the model's size and the card's memory bandwidth; real numbers depend on the runtime, drivers and context length. Community figures are medians of user reports.