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.

empero-ai logo

Can the Apple M4 (16 GB) run Qwen3.8-4B-Distill?

Qwen3.8-4B-Distill·4.7B·Apple M4 (16 GB)·16 GB unified memory

Saved in this browser, used on every model page.
Yes

Yes. The Apple M4 (16 GB) runs Qwen3.8-4B-Distill fully in unified memory at FP16 / BF16 (about 10.0 GB of 11 GB), at roughly 7.1 tokens/s.

best quant
FP16 / BF16
needs
10.0 GB
usable
11 GB
est. speed
~7.1 tok/s

Estimates for an 8K context; offload assumes 32 GB of system RAM.Try other contexts on the Qwen3.8-4B-Distill page →

Every quantisation of Qwen3.8-4B-Distill on the Apple M4 (16 GB)

QuantQualityMemoryFitSpeed
Q2_K Heavy loss3.4 GBfits~43 tok/s
Q3_K_M Noticeable loss3.8 GBfits~29 tok/s
Q4_K_M Small loss4.2 GBfits~23 tok/s
Q5_K_M Near-lossless4.6 GBfits~20 tok/s
Q6_K Near-lossless5.0 GBfits~17 tok/s
Q8_0 Lossless6.0 GBfits~13 tok/s
FP16 / BF16 pickLossless10.0 GBfits~7.1 tok/s

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 →

Run Qwen3.8-4B-Distill on the Apple M4 (16 GB)

Using FP16 / BF16, the best fit for Apple M4 (16 GB).

MLX · Apple Silicon with mlx-lmdocs
pip install mlx-lm
mlx_lm.generate --model yachen4ever/Qwen3.8-4B-Distill-Heretic-Abliterated-MLX-4bit --prompt "Hello"

Use `mlx_lm.chat` with the same --model for an interactive chat.

MLX · OpenAI-compatible serverdocs
mlx_lm.server --model yachen4ever/Qwen3.8-4B-Distill-Heretic-Abliterated-MLX-4bit --port 8080

API on http://localhost:8080/v1.

Ollama · BF16 straight from Hugging Facedocs
ollama run hf.co/empero-ai/Qwen3.8-4B-Distill-GGUF:BF16
llama.cpp · OpenAI-compatible server with a web UIdocs
llama-server -hf empero-ai/Qwen3.8-4B-Distill-GGUF:BF16

brew install llama.cpp ships with Metal. Downloads the GGUF on first run.

LM Studio · Download with the lms CLIdocs
lms get empero-ai/Qwen3.8-4B-Distill-GGUF@BF16

Then load it from the LM Studio app or with `lms load`.

Transformers · Chat with the pipeline APIdocs
from transformers import pipeline

pipe = pipeline("text-generation", model="empero-ai/Qwen3.8-4B-Distill", dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Explain self-hosting in one paragraph."}]
out = pipe(messages, max_new_tokens=512)
print(out[0]["generated_text"][-1]["content"])

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

All that fit →
ModelParamsBest quantMemorySpeedLicence
4BQ8_0≈5.8 GBFits in memory~15 tok/sCommercial OK
3.2BQ8_0≈4.7 GBFits in memory~19 tok/sConditions
4.1BQ8_0≈5.9 GBFits in memory~15 tok/sCommercial OK
3.8BQ8_0≈5.5 GBFits in memory~16 tok/sCommercial OK
4BQ8_0≈5.8 GBFits in memory~15 tok/sCommercial OK
8BQ8_0≈9.8 GBFits in memory~7.7 tok/sConditions

Other GPUs that run Qwen3.8-4B-Distill

Frequently asked questions

Can the Apple M4 (16 GB) run Qwen3.8-4B-Distill?
Yes. The Apple M4 (16 GB) runs Qwen3.8-4B-Distill fully in unified memory at FP16 / BF16 (about 10.0 GB of 11 GB), at roughly 7.1 tokens/s.
How fast does Qwen3.8-4B-Distill run on the Apple M4 (16 GB)?
An estimate from memory bandwidth gives about 7.1 tokens/s at FP16 / BF16.
Which quantisation of Qwen3.8-4B-Distill should I use on the Apple M4 (16 GB)?
FP16 / BF16: the largest one that fits. Expected quality: lossless 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.