Skip to content
mlx.app

Mixtral 8x7B Instruct v0.1

Mistral AI · 46.7B (12.9B active, MoE) · up to 32K context · apache-2.0

FITS · 11.1 GB FREE

One of the earliest MoE models to run well on Apple Silicon, needing a 32GB Mac for 4-bit despite having only 12.9B active parameters per token. It still runs faster than a dense 46B model would, but newer MoE designs like Qwen3-30B-A3B have mostly superseded it on quality.

On a M4 Pro with 48 GB

macOS9.6 GBweights26.3 GBKV cache1.1 GBheadroom11.1 GB
Weights
26.3 GB
KV cache
1.1 GB
Usable RAM
38.4 GB
Headroom
11.1 GB
Generation
29.3 tok/sest.
Prompt processing
70.0 tok/sest.
First token (1K prompt)
14.6 sest.

Decode speed is bandwidth ÷ active weight bytes, at 273 GB/s and 78% efficiency. These are modelled figures, not measurements.

Every quantization, on your Mac

QuantWeights+ KVVerdictSpeed est.Repo
q426.3 GB27.3 GBFITS · 11.1 GB FREE29.3 tok/sMixtral-8x7B-Instruct-v0.1-4bit
q849.6 GB50.7 GBOVER BY 12.3 GB15.5 tok/sMixtral-8x7B-Instruct-v0.1-hf-4bit-mlx

We never host weights. Every link goes to Hugging Face.

Run it

Chat
mlx_lm.chat --model mlx-community/Mixtral-8x7B-Instruct-v0.1-4bit
Serve an OpenAI-compatible endpoint
mlx_lm.server --model mlx-community/Mixtral-8x7B-Instruct-v0.1-4bit --port 8080

How far can you push the context

2K tokensFITS · 11.9 GB FREE

cache 0.3 GB

8K tokensFITS · 11.1 GB FREE

cache 1.1 GB

32K tokensFITS · 7.8 GB FREE

cache 4.3 GB

Which Macs run this

ChipSmallest RAM that fitsSpeed est.
M1no configuration
M1 Prono configuration
M1 Max64 GB43.5 tok/s
M1 Ultra64 GB89.3 tok/s
M2no configuration
M2 Prono configuration
M2 Max64 GB44.1 tok/s
M2 Ultra64 GB90.4 tok/s
M3no configuration
M3 Prono configuration
M3 Max (14-core CPU)no configuration
M3 Max (16-core CPU)48 GB44.1 tok/s
M3 Ultra96 GB93.7 tok/s
M4no configuration
M4 Pro48 GB29.3 tok/s
M4 Max (14-core CPU)48 GB45.8 tok/s
M4 Max (16-core CPU)48 GB61.7 tok/s
M5no configuration
M5 Prounverified48 GB33.4 tok/s
M5 Maxunverified48 GB70.2 tok/s