LLMLab.ee

Experimental Mac + eGPU

Mac mini M4 Pro 48GB + Radeon AI PRO R9700 eGPU Experiment

The external 32GB GPU memory and the Mac's 48GB unified memory are separate runtime pools. The exact ASUS card remains inside the Sonnet enclosure's published physical and power limits; that does not establish macOS software support.

experimental eGPU path

Is this build right for me?

AI strength score

The native Mac path sets this score; external-GPU compute remains a separate validation-only option.

Value

Weak local-AI capability for the current estimated price.

What this build can handle

Comfortable

  • Smaller native MLX or Ollama models that fit the Mac's shared memory.
  • Planning and reproducing the exact eGPU software stack before relying on it.

Possible with limits

  • 32 GB external-GPU inference after the exact driver and software environment are validated.
  • A monitored experimental workload after it is reproduced on this exact hardware.

Not recommended

  • Buying it for production work before the exact setup is qualified.
  • Official Apple Silicon eGPU support, macOS graphics or gaming acceleration, or plug-and-play native CUDA.

Price estimate history

Uses market prices where available and reference estimates for the rest.

Reference estimate

€7,102

System estimate

Time range

Tap a point to see its price. Swipe the chart sideways to inspect every date.

The chart slider reports the estimated system total by date.

Price estimate historyThe line shows the estimated system total over time. Values for individual dates remain available through the chart slider and data table.€6,210€6,575€6,941€7,306€7,67116 Jun31 Jul14 Sept

System estimate by date
DateSystem estimateMac systemeGPU enclosureGPU
16 Jun€6,901€2,753€806€1,733
21 Jun€6,901€2,753€806€1,733
26 Jun€6,901€2,753€806€1,733
1 Jul€6,901€2,753€806€1,733
6 Jul€6,901€2,753€806€1,733
11 Jul€6,901€2,753€806€1,733
16 Jul€6,901€2,753€806€1,733
21 Jul€6,901€2,753€806€1,733
26 Jul€6,901€2,753€806€1,733
31 Jul€6,901€2,753€806€1,733
5 Aug€6,901€2,753€806€1,733
10 Aug€6,901€2,753€806€1,733
15 Aug€6,901€2,753€806€1,733
20 Aug€6,901€2,753€806€1,733
25 Aug€6,901€2,753€806€1,733
30 Aug€6,901€2,753€806€1,733
4 Sept€7,102€2,753€806€1,927
9 Sept€7,102€2,753€806€1,927
14 Sept€7,102€2,753€806€1,927

Core Configuration

Mac System

Mac mini M4 Pro 48GB / 1TB

Chip

Apple M4 Pro

Unified Memory

48 GB

Storage

1000 GB

GPU

ASUS Turbo Radeon AI PRO R9700 32GB

VRAM

32 GB

Performance & Power

eGPU Enclosure

Sonnet Breakaway Box 850T5 Thunderbolt 5 eGPU

Architecture

RDNA 4

Risk level

Experimental

Workload fit

Mainly for 7B/8B models

Mac unified memory48 GBNative Mac path
External GPU VRAM32 GBQualified eGPU path

Qualification protocol

PRE-QUOTE
  1. 01Define one target workload
  2. 02Confirm enclosure and GPU path
  3. 03Reproduce driver and software stack
  4. 04Make the go/no-go decision after testing

Supported Workloads

  • The same explicitly qualified TinyGPU or tinygrad compute experiment as the main eGPU build
  • with more host-side unified memory and storage for normal MLX
  • development
  • and dataset work

Not Supported

  • Official Apple Silicon eGPU support
  • macOS graphics/display/gaming acceleration
  • Final Cut acceleration
  • native CUDA
  • general ROCm support on macOS
  • or production use without exact-stack qualification

Buyer Warning

UNQUALIFIED EXPERIMENTAL PATH: the larger Mac configuration does not make Apple Silicon eGPU support official. No hardware is quoted until the exact 48GB Mac, macOS, Sonnet GPU-850T5 enclosure, Radeon AI PRO R9700, TinyGPU/tinygrad version, PCIe mapping, compiler path, and target workload are reproduced successfully.

Component Pricing Breakdown

Component rows show planning prices for the Mac, eGPU enclosure, and GPU.

ComponentProductEstimated market price
Mac SystemMac mini M4 Pro 48GB / 1TB€2,891
eGPU EnclosureSonnet Breakaway Box 850T5 Thunderbolt 5 eGPU€846
GPUASUS Turbo Radeon AI PRO R9700 32GB€2,024
Estimated parts subtotal€5,761
Service, setup, and validation fee€1,623
Estimated total€7,384

Local AI examples

Entry local models onlyNot ideal for 70B+ models

Recommended model

Qwen3.6 27B

A strong dense model for 24GB to 32GB GPUs and higher-memory Apple systems.

Partial GPU offload only

Very slow

ollama run qwen3.6:27b

Experimental Mac + eGPU runtime support must be validated; treat this as a validation target, not a normal recommendation. Quantized GPU memory fit looks reasonable, but longer context can still add pressure.

  • 48GB system/unified memory available
  • 32GB effective accelerator memory for model weights and cache

Gemma 4 12B

Private assistant chat, document understanding, and image or audio analysis

Partial GPU offload only

A current multimodal everyday model that makes good use of a 16GB GPU.

Expected experience: Very slow

Experimental Mac + eGPU runtime support must be validated; treat this as a validation target, not a normal recommendation. Likely good memory headroom for this quantized model at normal context sizes.

  • 48GB system/unified memory available

Qwen3.6 35B-A3B

Coding agents, repository analysis, and complex local assistant workflows

Not realistic here

A capable MoE model that gives 32GB and 48GB workstations a meaningfully heavier coding target.

Expected experience: Not recommended

Needs at least 64GB system RAM; this machine reports 48GB.

  • 48GB system/unified memory available

Llama 3.3 70B Instruct

High-end local chat experiments on 48GB GPUs or 96GB+ Apple systems

Not realistic here

A clear upper-limit example for checking whether a machine can attempt a 70B-class model.

Expected experience: Not recommended

Needs at least 64GB system RAM; this machine reports 48GB.

  • 48GB system/unified memory available
Expandable technical details

Assumptions

  • GPU VRAM assumption: 32GB from ASUS Turbo Radeon AI PRO R9700 32GB.
  • System RAM: 48GB.
  • Experimental Mac + eGPU fit depends on driver/runtime support, not just VRAM.
  • Ratings include model weights, estimated KV cache, runtime overhead, and safety margin for one local model running at a time. Treat them as fit guidance, not a speed guarantee.
  • This setup depends on experimental Mac + external GPU software support. Treat compatibility ratings as a starting point for the pre-quote consultation.
Qwen3.6 27B technical details

Family: Qwen3.6

Parameters: 27B

Structure: 27B dense

License: Apache License 2.0; allowed. For guidance only; review the model licence before commercial use.

Native context: 262,144 tokens

Extended context: 1,010,000 via YaRN; not a default fit assumption

Quantization: Q4_K_M

Approx. Q4 weights: 17 GB

Default estimate: 23 GB @ 8,192 tokens

Weights / KV / runtime / margin: 17 GB / 2 GB / 1.5 GB / 2.5 GB

CPU/RAM fallback: Not recommended

VRAM: 24 GB minimum / 32 GB recommended

RAM: 48 GB minimum / 64 GB recommended

Current run mode: Partial GPU offload only

Expected experience: Very slow

Full GPU offload: Only when the memory estimate and context fit

Context warning: Repository-scale or very long document context can push a 24GB card beyond a comfortable fit.

ContextWeightsKVRuntimeMarginEstimated GPU memory
4K17 GB1 GB1.5 GB2.5 GB22 GB
8K17 GB2 GB1.5 GB2.5 GB23 GB
16K17 GB4 GB1.5 GB3 GB25.5 GB
32K17 GB8 GB1.5 GB3.5 GB30 GB

Swipe the table sideways to inspect every memory estimate.

Research sources

Qwen3.6 27B official model cardOllama Qwen3.6 tags

Researched: 2026-07-30

Gemma 4 12B technical details

Family: Google Gemma 4

Parameters: 11.95B

Structure: 11.95B dense

License: Apache License 2.0; allowed. For guidance only; review the model licence before commercial use.

Native context: 262,144 tokens

Quantization: Q4_0 QAT

Approx. Q4 weights: 7.2 GB

Default estimate: 11.5 GB @ 8,192 tokens

Weights / KV / runtime / margin: 7.2 GB / 1.5 GB / 1.2 GB / 1.5 GB

CPU/RAM fallback: Not recommended

VRAM: 12 GB minimum / 16 GB recommended

RAM: 24 GB minimum / 32 GB recommended

Current run mode: Partial GPU offload only

Expected experience: Very slow

Full GPU offload: Only when the memory estimate and context fit

Context warning: Its large context window still requires substantial KV-cache headroom.

ContextWeightsKVRuntimeMarginEstimated GPU memory
4K7.2 GB1 GB1.2 GB1.5 GB11 GB
8K7.2 GB1.5 GB1.2 GB1.5 GB11.5 GB
16K7.2 GB2.5 GB1.2 GB1.5 GB12.5 GB
32K7.2 GB5 GB1.2 GB2 GB15.5 GB

Swipe the table sideways to inspect every memory estimate.

Research sources

Google Gemma model overviewGemma 4 release notesOllama Gemma 4 tags

Researched: 2026-07-30

Qwen3.6 35B-A3B technical details

Family: Qwen3.6

Parameters: 35B

Structure: 35B total / 3B active MoE

License: Apache License 2.0; allowed. For guidance only; review the model licence before commercial use.

Native context: 262,144 tokens

Extended context: 1,010,000 via YaRN; not a default fit assumption

Quantization: Q4_K_M

Approx. Q4 weights: 24 GB

Default estimate: 31.5 GB @ 8,192 tokens

Weights / KV / runtime / margin: 24 GB / 2.5 GB / 2 GB / 3 GB

CPU/RAM fallback: Not recommended

VRAM: 32 GB minimum / 48 GB recommended

RAM: 64 GB minimum / 64 GB recommended

Current run mode: Not realistic here

Expected experience: Not recommended

Full GPU offload: Only when the memory estimate and context fit

Context warning: Long repository context can use the remaining headroom quickly, especially on 32GB GPUs.

ContextWeightsKVRuntimeMarginEstimated GPU memory
4K24 GB1.5 GB2 GB3 GB30.5 GB
8K24 GB2.5 GB2 GB3 GB31.5 GB
16K24 GB5 GB2 GB3.5 GB34.5 GB
32K24 GB10 GB2 GB4 GB40 GB

Swipe the table sideways to inspect every memory estimate.

Research sources

Qwen3.6 35B-A3B official model cardOllama Qwen3.6 tags

Researched: 2026-07-30

Llama 3.3 70B Instruct technical details

Family: Meta Llama 3.3

Parameters: 70.60B

Structure: 70.6B dense

License: Llama 3.3 Community License; allowed with terms; gated access. For guidance only; review the model licence before commercial use.

Native context: 131,072 tokens

Quantization: Q4_K_M

Approx. Q4 weights: 42.5 GB

Default estimate: 51 GB @ 4,096 tokens

Weights / KV / runtime / margin: 42.5 GB / 3 GB / 1.5 GB / 4 GB

CPU/RAM fallback: Not recommended

VRAM: 48 GB minimum / 64 GB recommended

RAM: 64 GB minimum / 96 GB recommended

Current run mode: Not realistic here

Expected experience: Not recommended

Full GPU offload: Often limited

Context warning: The Q4 weights alone use about 42.5GB before KV cache, runtime overhead, and desktop headroom.

ContextWeightsKVRuntimeMarginEstimated GPU memory
4K42.5 GB3 GB1.5 GB4 GB51 GB
8K42.5 GB6 GB1.5 GB4 GB54 GB
16K42.5 GB12 GB1.5 GB4.5 GB60.5 GB
32K42.5 GB24 GB1.5 GB5.5 GB73.5 GB

Swipe the table sideways to inspect every memory estimate.

Research sources

Meta Llama 3.3 model cardOllama Llama 3.3 70B tag

Researched: 2026-07-30

Local AI performance is approximate. Results depend on quantization, context length, backend, drivers, and whether the model plus KV cache fits in VRAM or Apple unified memory.

AI terms in plain language

GPU memory and RAM

GPU memory usually limits model size; RAM supports apps, data, and model offload.

CUDA

NVIDIA’s software layer for many AI tools. Macs and AMD GPUs do not run CUDA workflows the same way.

Apple unified memory

Memory shared by the CPU, GPU, macOS, and apps; it is not the same as NVIDIA VRAM.

7B / 8B / 14B / 70B

Approximate model size in billions of parameters; larger models usually need more memory.

Quantized models

Lower-precision models, such as Q4, that use less memory with possible quality or speed tradeoffs.

Context length

How much text the model can keep in mind at once. Longer context uses more memory.

Inference

Running an existing model for chat, coding help, summaries, or document workflows.

Fine-tuning vs adapter tuning

LoRA and QLoRA train small adapters and need fewer resources than full fine-tuning.

Throughput (tokens/s)

Model output speed. Compare the same model, quantization, context, runtime, and user count; prompt processing is a separate measurement.

Request a quote

€7,384

Estimated total. Final price confirmed before ordering.

Request a quote
Back to catalog