By LLM Hardware team · Methodology ·
Model fit and Mac prices rechecked 13 October 2026
Checked 2026-10-13. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Key fact: A 24 GB Mac is tight for Qwen3.8-27B at Q4. A 36 GB Studio is the comfortable Apple fit. LFM2.5-24B-A2B Q4 is 16.4 GB and is tight on 24 GB unified.
How we got it: 27 × 0.5 × 1.2 + 2 = 18.2 GB. 24 × 0.5 × 1.2 + 2 = 16.4 GB. Subtract about 4 GB for macOS before judging the fit. Check current price on Amazon.
Checked:
Key fact: Gemma 4 26B-A4B needs 17.1 GB at Q4 when counted at 25.2B total parameters.
How we got it: 25.2 × 0.5 × 1.2 + 2. Google card: 3.8B active does not shrink the resident experts. Unsloth’s 4-bit note is 16–18 GB.
Checked:
Methodology — parameter math, quantization bytes, and the source list.
Unified memory is still the reason to buy a Mac for local models: the GPU can use the same pool as the CPU, so capacity matters more than a discrete VRAM ceiling. What changed by 1 October 2026 is which boxes Amazon actually had in a clean buy box. The in-stock Apple path tops out at 36GB. Anything at 64GB or above — including M5 Ultra — is an Apple Store configure, and this guide does not invent an ASIN for it.
Apple's current Mac mini comes with two chips. The M6 starts at 16 GB of unified memory and can be ordered with 24 GB or 32 GB. It is the same chip at every size, but the bandwidth changes with the memory: Apple lists 153 GB/s for the 16 GB model and 170 GB/s for the 24 GB and 32 GB models. The M5 Pro starts at 24 GB and goes to 48 GB or 64 GB, at 307 GB/s. The M4 models are now the previous generation.
For local models, the memory tier decides what fits. We hold back 4 GB for macOS, so a 16 GB M6 leaves about 12 GB for the model and a 32 GB M6 leaves about 28 GB. By our formula, a 14B model at Q4 needs about 10.4 GB, a 27B about 18.2 GB, a 32B about 21.2 GB and a 70B about 44 GB. That puts a 14B on the 16 GB M6, and a 27B at Q4 on the 32 GB M6 with room left for context. A 70B at Q4 would use all of a 48 GB M5 Pro's usable memory, so it needs a 64 GB M5 Pro or a Mac Studio.
The 16 GB and 24 GB M6 models are sold on Amazon. The 32 GB M6, the 48 GB and 64 GB M5 Pro and the Mac Studio builds are ordered from Apple.
In stock, sold by Amazon.com (2026-10-18).
Model fit. After a 4 GB macOS reserve, about 12 GB is usable. 7B Q4 is 6.2 GB and fits. 8B Q4 is 6.8 GB and fits. Gemma 3 12B Q4 is 9.2 GB and fits with little room. 14B Q4 is 10.4 GB and is tight. 27B Q4 is 18.2 GB and does not fit.
Check current price on AmazonSold by Amazon.com. Usually ships within 3 to 5 weeks (2026-10-18).
Model fit. After a 4 GB macOS reserve, about 20 GB is usable. 14B Q4 is 10.4 GB and fits. 27B Q4 is 18.2 GB and is tight. 32B Q4 is 21.2 GB and does not fit.
Check current price on AmazonIn stock, Amazon.com (2026-10-17).
Model fit. After a 4 GB macOS reserve, about 32 GB is usable. Qwen3.6-27B Q4 is 18.2 GB and fits. LFM2.5-24B-A2B Q4 is 16.4 GB and fits. Q8 is 30.8 GB and is the edge. 70B Q4 is 44 GB and does not fit.
Check current price on AmazonIn stock, Amazon.com (2026-10-17).
Model fit. After a 4 GB macOS reserve, about 20 GB is usable. 27B Q4 is 18.2 GB and is tight. LFM2.5-24B-A2B Q4 is 16.4 GB and is tight. Q8 is 30.8 GB and does not fit.
Check current price on AmazonIn stock (2026-10-17).
Model fit. 128 GB at ~256 GB/s. Dense 70B Q4 (44 GB) fits and is slow. Qwen3.8-Flash-Next UD-IQ4_XS is 93.68 GB and fits. UD-Q4_K_XL is 111.33 GB and is tight. DeepSeek-V4-Flash only at an aggressive quant (~155 GB at a normal 4-bit). Not CUDA.
Check current price on AmazonCatalog entries live on the hardware index. Compare any two of them on the compare page. Check a specific model’s footprint with the VRAM calculator.
The names in this table were mapped on 1 October and the Qwen3.8 / Gemma 4 cells were recalculated on 3 October with our VRAM formula. DeepSeek-V4-Flash, GLM-5.3-Flash, and the other large MoEs stay in the 96 GB-and-up rows. They are not 24 GB models. No new llama-bench run was done for this pass.
| Memory tier | Hardware in this check | Example models |
|---|---|---|
| 6–12 GB usable | Mac mini M6 16GB. M4-era 16GB minis are the previous generation and are not a buy pick. | K2-Horizon-0.9B, 3.7B, and 7B at Q4–Q5. Leave about half the 16GB pool for the OS. |
| 16–24 GB | Mac mini M6 24GB and Mac mini M5 Pro 24GB. Mac Studio M5 Max 36GB is the top of this band. | Qwen3.8-27B Q4 is 18.2 GB by the formula (Ollama listed 18 GB). Tight on 24 GB unified, comfortable on 36 GB. Gemma 4 26B-A4B is 17.1 GB on 25.2B total, same split. Gemma 4 E2B/E4B fit the smaller end of this band. |
| 48–64 GB | M5 Pro 48GB and 64GB, and Mac Studio M5 Max, ordered from Apple. M4-era machines in this band are the previous generation and are not a buy pick. | Qwen3.8-27B at higher precision, K2-Horizon-32B with room, and small MoE experiments. 70B-class MoE only at aggressive quants. |
| 96–128 GB | GMKtec EVO-X2 128GB. Mac Studio M5 Ultra is ordered from Apple. | Qwen3.8-Flash-Next UD-IQ4_XS is 93.68 GB and fits 128 GB. It does not fit the 96 GB M5 Ultra base once macOS is reserved. UD-Q4_K_XL is 111.33 GB and is tight on 128 GB and does not fit 96 GB. A 70B Q4 file is 44 GB and fits the 96 GB Ultra. DeepSeek-V4-Flash at a normal 4-bit is about 155 GB and does not fit either box. |
| 192 GB+ | Apple Ultra 256–512GB on the Apple configure page. Framework Desktop 192GB showed up in community discussion and was not Amazon-verified in this Mac pass. | DeepSeek-V4.1-Flash at low bits, GLM-5.3-Flash, MiniMax-M3, and larger MoEs. |
If the model you care about is a large MoE, the in-stock Studio is the wrong box. These configs were on Apple’s buy page and did not have a US Amazon ASIN we could verify. An EU tracker ASIN (B0HGRN2PV2) returned 404 on amazon.com, so it is not listed.
Start at apple.com/shop/buy-mac/mac-studio. The 36GB M5 Max base is also sold on Amazon; the higher memory steps are the Apple-only part.
Community discussion in September and October 2026 kept pointing at high unified memory — Studio M5 Max and Ultra, Strix Halo 128GB, Framework Desktop 192GB — for Qwen3.8, Flash-Next, and Muse. The default Amazon PC in that class is the GMKtec EVO-X2: 128 GB at about 256 GB/s. The Beelink GTR9 Pro is the same APU and bandwidth with a higher price and a 10GbE NIC history; it is a caveat on the Strix Halo guide, not a second default. Plan on Linux or a ROCm/Vulkan stack. CUDA workflows still belong on an NVIDIA card; see the GPU buying guide and the Mac vs PC guide.
DeepSeek’s current big local target in this window is DeepSeek-V4-Flash (~155GB at a normal 4-bit). That is an EVO-X2 aggressive quant or an Apple Ultra, not a 24GB mini. Qwen sizing for the dense 27B class is covered in the Qwen hardware guide.
On the Macs, Ollama is the always-on worker (brew install ollama, then a model pull). For the 3 October models the tags are ollama run qwen3.8:27b and ollama run gemma4:26b on a 36 GB Studio, and gemma4:e4b or gemma4:e2b when the pool is smaller. There is also a qwen3.8:27b-mlx tag if you want the MLX build Ollama listed. LM Studio is the GUI. llama.cpp with Metal is the same engine with the flags exposed. MLX (pip install mlx-lm) is Apple’s framework and the one to reach for when you want to fine-tune on the GPU you already bought. Setup steps are in How to run LLMs on Mac and the Ollama cheat sheet. Check the number first in the VRAM calculator. The quantization trade itself is in the quantization guide and how much VRAM you need. The tier that says “does not fit 16 GB” is on what you can run.
A Mac mini in this tier is a reasonable always-on machine for agents and a self-hosted runner. It is not a substitute for 64GB+ when the weights do not fit. Memory is soldered. Buy the pool you need on day one.
Qwen3.6-27B Q4 is 18.2 GB: tight on 24 GB unified, comfortable on the 36 GB Studio. LFM2.5-24B-A2B Q4 is 16.4 GB: tight on 24 GB unified after a ~4 GB reserve, and it fits the 36 GB Studio. Q8 is 30.8 GB and is the edge of 36 GB after that reserve. Qwen3.8-Flash-Next UD-Q4_K_XL is 111.33 GB and fits neither Mac. The M5 Ultra 96GB base is an Apple configure with no verified US dollar and no Amazon ASIN. After a 4 GB macOS reserve it holds a 70B Q4 file (44 GB) and does not hold Flash-Next (110 GB). Check current price on Amazon.
Our VRAM formula is params × bytes × 1.2 + 2. Qwen3.8-27B at Q4 is 18.2 GB, and Ollama listed an 18 GB download on 3 October 2026. That fits a 24 GB discrete GPU at a short context and is tight on a 24 GB Mac because macOS shares the pool. A 36 GB Mac Studio M5 Max is the comfortable Mac for that model and for Gemma 4 26B-A4B (17.1 GB on its 25.2B total). Skip 32B Q8 (40.4 GB) on both. It is not a DeepSeek-V4-Flash machine, and it is not a Kolibri machine.
Not as a verified in-stock Amazon buy on 2026-10-01. The Mac Studio M4 Max 64GB/1TB listing (ASIN B0FNS1ZX5B) was currently unavailable. 48GB Mac mini M4 Pro listings (B0DS2XP86K and B0DYFCQ7CB) were also unavailable. M5 Max 64GB and M5 Ultra 96GB are configured at apple.com/shop/buy-mac/mac-studio. This guide does not invent Amazon ASINs for those configs.
It is the verified Amazon PC alternative when unified memory above 64GB is the goal. The EVO-X2 (Ryzen AI Max+ 395, Strix Halo) has 128GB soldered LPDDR5X at about 256 GB/s. Qwen3.8-Flash-Next UD-IQ4_XS is 93.68 GB and fits. UD-Q4_K_XL is 111.33 GB and is tight. A dense 70B Q4 is 44 GB and is slow at this bandwidth. The old GTR9 ASIN B0GQXDCKN1 loaded a Beelink SER3 that day, so it is not a buy link. The EVO-X2 is not CUDA. Check current price on Amazon.
Gemma 4 E2B and E4B fit smaller Macs. Unsloth’s Q4_K_M files were 3.11 GB and 4.98 GB on 3 October 2026. Gemma 4 26B-A4B is 25.2B total and 3.8B active, Apache-2.0, and the Q4 formula on the total is 17.1 GB, so it follows the same 24 GB tight / 36 GB comfortable split as Qwen3.8-27B. Kolibri-1 does not change a consumer Mac purchase. Aleph Alpha’s card lists about 78 GB of FP8 weights and a minimum of datacenter GPUs (two A100 80 GB or two H100s, or one H200, B200, or B300). That is not a 24 GB mini and not a claim that a 128 GB box matches their tested stack.
Ollama, LM Studio, llama.cpp with Metal, and Apple MLX all run on Apple Silicon. Ollama is the usual always-on worker. MLX is the path when you want Apple’s own framework, including fine-tuning experiments. None of these replace buying enough unified memory for the model.
Check a model against these machines, or compare them side by side.
How to run LLMs on Mac
Ollama, LM Studio, llama.cpp, and MLX on Apple Silicon.
Mac vs PC for LLMs
When unified memory wins, and when a discrete GPU is the faster buy.
Best GPU for LLMs
NVIDIA and AMD picks when CUDA or raw tok/s matters more than a Mac.
Qwen hardware requirements
Where the dense Qwen sizes land, including the 27B class.
DeepSeek hardware requirements
Why V4-Flash is a 96GB+ problem, not a mini problem.
How much VRAM do I need?
The same Q4 formula, worked through for discrete cards.
Quantization explained
Why Q4 is the planning quant, and where quality drops.
Beginner’s guide to local AI
If you have not installed a local model yet, start here.
A number that does not match a listing? Contact us.