Can the Apple M3 run Qwen3-8B?

Qwen3-8B · Apple M3 · 8,192 ctx · whole model residentFP16 sizereconstructed sizeSpeedestimate

IT FITS

Yes. 9 of 9 formats evaluated fit in 24 GB, the largest Apple M3 configuration, at 8,192 tokens. 5 of them are sized from a published file; the rest are rebuilt from the pinned architecture. The best quality that fits is FP16, needing 18.4 GB and running at an estimated 6 tokens per second.

Fitting in memory is not the same as loading. Whether the runtime and version you have supports this architecture and format on this machine has not been tested here.

Find the Apple M3: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

9/9formats that fit

24 GBdevice memory, largest configuration

18.4 GBneeded at best quality

~5.6tokens per second, estimatedSlow

100GB/s bandwidth

Run it in the cloud

Run Qwen3-8B properly, for cents an hour

The Apple M3 only holds it at FP16, at an estimated ~5.6 tokens per second. A rented card runs a better format at full speed, billed by the second.

Prices · 18:30 UTC, 22 Sept

Recommended · Cheapest12 GB

1× RTX 3060

$0.036per hour

Speed: ~41 tok/sFaster than you read
Per million tokens: $0.24

Vast.ai · Marketplace · 99.8% reliable

Rent on Vast.ai

Lowest cost per token at 20+ tok/s

2.1× faster10 GB

1× RTX 3080

$0.082per hour

Speed: ~87 tok/sFaster than you read
Per million tokens: $0.26

Vast.ai · Marketplace · 98.3% reliable

Rent on Vast.ai

Cheapest at 1.6× the speed or more

Fastest270 GB

1× B300

$7.89per hour

Speed: ~889 tok/sFaster than you read
Per million tokens: $2.46

RunPod · Secure Cloud

Rent on RunPod

Highest speed estimate

Compare all 28 rentable configurations →

No GPU at all

Use Qwen3-8B by the token

OpenRouter$0.46 / M output · $0.12 / M inputOpen on OpenRouter

Here the rented card is as cheap as the API or cheaper per token ($0.24 per million), and it keeps your data on a machine you control. The API is still the easier start: nothing to set up and nothing to switch off.

First time renting a GPU? How it works, in four steps

Create an account and add credit. Both providers are prepaid: $10 is enough to try any card on this page for hours. A new RunPod account that arrives through a referral link gets a one-time $5 credit when it first adds $10.
Pick the card and the count shown here, and a template: llama.cpp or vLLM for an OpenAI-compatible API, or one with a web chat if you only want to talk to the model. “How to launch it” below gives the exact image and command.
Wait for the download. The weights are fetched on the machine; tens of gigabytes take a few minutes on a data-centre connection. You pay by the second from the moment it starts.
Stop it when you are done. A running machine bills even when idle. A stopped one costs nothing per hour, though its disk is usually still billed until you delete it.

How to launch Q4_K_M on it (llama.cpp)

Container image: ghcr.io/ggml-org/llama.cpp:server-cuda
Start command / arguments: -hf Qwen/Qwen3-8B-GGUF:Q4_K_M -c 8192 -ngl 999 --host 0.0.0.0 --port 8080
Expose port: 8080 — an OpenAI-compatible API at /v1

Starting from a saved llama.cpp template instead? Leave the arguments empty and set these variables — only the first two change between models:

LLAMA_ARG_HF_REPO=Qwen/Qwen3-8B-GGUF:Q4_K_M
LLAMA_ARG_CTX_SIZE=8192
LLAMA_ARG_N_GPU_LAYERS=999
LLAMA_ARG_HOST=0.0.0.0
LLAMA_ARG_PORT=8080

Already have the runtime? llama-server -hf Qwen/Qwen3-8B-GGUF:Q4_K_M -c 8192 -ngl 999 --host 0.0.0.0 --port 8080. Weights: Qwen/Qwen3-8B-GGUF. The CUDA image is for NVIDIA cards; an AMD machine needs llama.cpp's ROCm build.

Every machine holds the whole model at Q4_K_M and 8,192 tokens of context, no offload. Speeds are this site’s single-stream decode estimates; prices are what each provider’s own API quoted, on-demand, for the whole machine. Vast hosts below 98% measured reliability are left out.

Referral links Vast.ai, RunPod and Novita pay us a share of what you spend if you sign up through these buttons. It costs you nothing, and it never decides an order or a recommendation: both are computed from the live price and the speed, and options that pay us nothing are listed and recommended on the same terms. How we rank

Runs well on your Apple M3 instead

The models nearest to Qwen3-8B — same lab first, then closest in size — that the Apple M3 holds whole at the standard format and runs at a usable speed. Nearest in size is not the same as equally good; compare them on the task you care about.

Qwen2.5-VL-7B-InstructQ4_K_M · 6.36 GB · ~17 tok/s
Qwen2.5-Coder-7B-InstructQ4_K_M · 5.95 GB · ~17 tok/s
Qwen2.5-7B-InstructQ4_K_M · 5.95 GB · ~17 tok/s
Qwen3.5-4BQ4_K_M · 3.98 GB · ~29 tok/s

How long a conversation. At FP16, this pair holds 40,960 tokens of context — roughly 30,720 words — needing 23.2 GB. That is Qwen's own configured maximum of 40,960 tokens, not the card running out. A larger device does not extend it; a documented RoPE or YaRN extension might, and is a separate question from this one. The nearest round setting below it is 32,768.

By memory configuration

The Apple M3 is sold with 8, 16 or 24 GB of unified memory, and the amount decides the fit. Each row is the best format that fits in that much at 8,192 tokens, whole model resident.

Memory	Best format that fits	Needs	Calculator
8 GB	Q5_K_M	7.86 GB	Open →
16 GB	Q8_0	10.7 GB	Open →
24 GB	FP16	18.4 GB	Open →

Every format evaluated

5 of them are sized from a published file; the rest are rebuilt from the pinned architecture. The Basis column says which is which for each row. Decode and first-token figures are calibrated estimates, not runs on this card.

Format	Needs	Of 24 GB	Fits	Decode	First token	Basis
FP16	18.4 GB	77%	yes	~5.6 tok/sSlow	no comparable peak	reconstructed size
Q8_0	10.7 GB	45%	yes	~9.5 tok/sSlow	no comparable peak	published data
Q6_K	8.73 GB	36%	yes	~12 tok/sAbout reading pace	no comparable peak	published data
Q5_K_M	7.86 GB	33%	yes	~13 tok/sAbout reading pace	no comparable peak	published data
Q5_0	7.73 GB	32%	yes	~13 tok/sAbout reading pace	no comparable peak	published data
Q4_K_M	7.04 GB	29%	yes	~15 tok/sAbout reading pace	no comparable peak	published data
Q4_0	6.78 GB	28%	yes	~15 tok/sAbout reading pace	no comparable peak	reconstructed size
Q3_K_M	6.16 GB	26%	yes	~17 tok/sAbout reading pace	no comparable peak	reconstructed size
Q2_K	5.21 GB	22%	yes	~20 tok/sAbout reading pace	no comparable peak	reconstructed size

Sized at 8,192 tokens of context with the whole model resident — weights, the KV cache and the runtime reserve, offload off. Speed is only quoted for a format that fits: a rate for a configuration that cannot load is not a fact about anything.

Why the first-token figure is the same on every row

First token is modelled from the arithmetic the prompt requires, and that count does not change with the weight format — which is why it reads the same on every row. Real prefill does vary by format, because a quantized matmul is a different kernel; this model does not capture that, and the figure should be read as an order of magnitude rather than a ranking between formats.

What the labels mean

published data: Read from a published source — a file's byte count, a model's configuration or a manufacturer's specification — or exact arithmetic on such values. Not a measurement on a machine.
reconstructed size: Weight size reconstructed from the pinned architecture, because no published file exists.
size range: Only a lower and an upper bound are claimed for this weight size.
estimate: Calculated from sourced inputs by a stated method; an estimate, not a measurement.

Change anything

This page fixes the context at 8,192 tokens and one device. Batch, concurrent users, KV-cache format, clusters and rental cost are all in the calculator, already set to this pairing.

Open in the calculator →Everything the Apple M3 runs Best models for 24 GB

Where these numbers come from

Qwen3-8B’s architecture is read from its publisher’s own config.json at a pinned revision, and the Apple M3’s 24 GB and 100 GB/s come from the manufacturer’s specification. How far each figure can be trusted is published, per format and worst case included, on the accuracy scorecard.

Building this into your own product? A free API key returns exactly these figures, and the widget puts this answer on a product page with one script tag.

Use this answer in your own product

As JSON, from the API

curl -s https://llmbottleneck.com/v1/analyze \
  -H "Authorization: Bearer $LLMB_KEY" -H "content-type: application/json" \
  -d '{"model":"qwen-qwen3-8b","quantization":"FP16","context":8192,"hardware":"apple-m3"}'

Same engine, same evidence, every field sourced. Free key in one step →

On your page, as a widget

<script src="https://llmbottleneck.com/widget.js"
  data-model="qwen-qwen3-8b" data-quantization="FP16"
  data-hardware="apple-m3" data-context="8192"></script>

No key needed. Unbranded, with your own buy button, on Pro and Business →