Splash Engine Mac Requirements 2026: The 36GB Floor That Rules Out the $899 M6 Mac mini

Inco AI shipped Splash on 17 September 2026 — a Qwen-specific inference engine for Apple Silicon — and LM Studio integrated it the next day. The decode numbers are genuinely good: roughly two to three times the next-fastest engine on the same machine. Every post currently ranking for it explains what Splash is.

None of them answer the question a buyer actually has, so here it is up front:

Splash Engine requires at least 36GB of unified memory, which means the $899 M6 Mac mini — capped at 32GB — can never run it, and the $1,699 M5 Pro Mac mini needs a $600 upgrade to its 48GB tier before it can, putting the cheapest Splash-capable Mac mini at $2,299.

Both machines went on sale six days ago, on 22 September 2026. The M6 mini is the cheapest new Mac and the one most people are about to buy. If you opened this page with an apple.com cart in another tab, that sentence is the whole answer: $899 does not buy Splash, and no amount of configuring makes it.

What follows is the eligibility matrix by SKU, what the speed actually buys you, the two-model catch the vendor posts bury, and the honest cross-shop against a 24GB or 32GB NVIDIA card.

The hardware gate, stated once

From Inco AI's launch post, verbatim: "Splash needs an M3 or newer Mac on macOS 26.4 or later with at least 36 GB of unified memory, and Homebrew. We recommend 48 GB or more." LM Studio restates the same floor in its own integration announcement.

Requirements from Inco AI's "Splash: A Local Engine Built Around the Model" (17 September 2026) and the LM Studio Splash Engine announcement (18 September 2026). Both pages opened and read 28 September 2026.

Now the verdict per shipping SKU. Memory tiers below come from Apple Newsroom and MacRumors' M6 vs. M5 Pro buyer's guide — no estimates, no interpolation.

Two of the four machines Apple currently sells as "Mac mini" configurations fail. That is the story.

Why the $899 M6 Mac mini can't run it, and won't later

The 32GB cap on the M6 is not a software restriction or a marketing tier. Unified memory on Apple Silicon is packaged with the SoC; the M6's maximum configuration is 32GB and that is the end of the conversation. There is no memory slot, no eGPU memory pool, no swap trick, and no macOS update that will conjure 4 more gigabytes.

This deserves saying carefully, because the M6 is a genuinely strong machine for local AI and conflating the two questions is how readers get mis-sold. Apple's own figure, verbatim from its Newsroom announcement: "Up to 13.5x faster LLM prompt processing in LM Studio when compared to Mac mini with M1, and up to 4.8x faster than M4." That is real, it is large, and it applies to the engines that ~99% of local-AI users actually run.

So separate the two questions, because every other post on this topic collapses them:

• "Can the M6 mini run local LLMs?" — Yes, well. MLX, llama.cpp, and Ollama all run on it, and a 27B-class model at Q4 fits inside 32GB with usable context. See our MLX vs llama.cpp comparison for which runtime to pick.
• "Can the M6 mini run Splash?" — No. Permanently no.

There is one loose thread worth naming rather than quietly using. A YouTube hands-on has been circulating with figures of 26–56 tok/s "on Splash" from a 32GB M6 mini, reached only through an aggregator's written summary of the video. If that number were real on a 32GB machine it would contradict both vendor pages and invert this entire post. We are not treating it as data: the aggregator does not state that Splash executed on the M6, and a video summarised by a third party is not a source we will publish against. Both first-party pages say 36GB. We go with 36GB.

A practical wrinkle that makes the floor tighter than it looks: macOS does not hand an application all of your unified memory. The default wired-memory limit for GPU work is roughly three-quarters of installed RAM, so a 36GB machine is working with something closer to 27GB of practical budget — which is precisely why Inco recommends 48GB rather than the bare 36GB minimum. Our memory-fit tool models that headroom rule against real model sizes.

The 24GB trap on the M5 Pro Mac mini

The M5 Pro Mac mini is the cheapest Splash-eligible Mac — but not at the price you see advertised. The $1,699 configuration ships with 24GB of unified memory. It is 12GB short of the floor, and because Apple Silicon memory is soldered, there is no remedy after the box arrives.

Apple offers three tiers on this machine: 24GB, 48GB, and 64GB. Splash starts at the middle one. So the buying sequence is:

1. Configure the M5 Pro Mac mini on Apple's store — do not add it to the cart at the default spec.
2. Change unified memory from 24GB to 48GB. This is the step that makes the machine eligible; it also happens to be Inco's recommended tier rather than the bare minimum.
3. Storage is unrelated to Splash eligibility — 512GB is fine unless you're hoarding model weights.

That step costs $600. Macworld's M5 Pro Mac mini review itemises its $2,899 review unit as the $1,699 base plus the full M5 Pro chip (+$200), 48GB of memory (+$600), a 1TB SSD (+$300), and 10 gigabit Ethernet (+$100) — an itemisation that reconciles exactly to its stated total. The chip, storage, and Ethernet upgrades are all optional and none of them affect Splash eligibility.

So the real number: a Splash-eligible Mac mini is $2,299 — the $1,699 base plus the $600 memory upgrade, nothing else. That is the price to compare against, not $1,699.

And it reframes the next machine up. A Mac Studio M5 Max starts at $2,499 and is Splash-eligible at its 36GB base — $200 more than the Splash-capable mini, for a considerably stronger GPU, 614GB/s of bandwidth against 307GB/s, and more ports. The mini wins on memory at that price (48GB vs 36GB) and on size; the Studio wins on almost everything else. If you were going to configure the mini up to 48GB anyway, price the Studio before you click buy.

The point underneath the figures: the "from $1,699" price is not the Splash-capable price. Our catalog entry for this machine lists "From $1,699" with a base 24GB configuration, and that is accurate — it is simply not the config this post recommends.

Apple's claim for the M5 Pro mini, from the same Newsroom announcement, for completeness: "Up to 8.5x faster LLM prompt processing performance in LM Studio when compared to Mac mini with M2 Pro, and up to 4x faster than M4 Pro." Combined with 307GB/s of bandwidth against the M6's 170GB/s, the M5 Pro is the better local-AI machine on the merits even before Splash enters the picture.

What the speed actually buys you (measured, sourced, and labelled)

Here are Inco's published figures, on a 48GB M5 Pro, by context length. Read the context column carefully — collapsing a "74 tok/s" headline into a general claim is how these numbers get abused.

Vendor-reported, not independently verified. Every figure above is Inco's own measurement of Inco's own engine, published at inco.ai/blog/splash. LM Studio's integration post restates the 74 / 54 tok/s pair and adds 170 tok/s combined across four concurrent requests on the 27B — LM Studio is the distributor, not a third party. No independent benchmark of Splash exists as of 28 September 2026.

Two things a shopper needs to understand before reading those rows as a hardware verdict:

The 35B-A3B is a Mixture-of-Experts model with roughly 3B active parameters per token. That is why a nominally larger model triples the throughput of the dense 27B — it is doing far less arithmetic per token while still occupying memory for all 35B of weights. If you shop on the 210 tok/s number expecting 27B-class output quality, you have mis-shopped. Our guide to running large MoE models on small GPUs covers that memory-versus-compute split, and the Qwen3.6 hardware guide covers the family this model comes from.

The dense 27B's decode rate barely degrades with context — 74 short, then a flat 54–55 from 8K all the way to 32K. That flatness is the genuinely impressive engineering result here, and it comes from aggressive KV cache handling. For anything agentic, where context grows monotonically across a session, flat-with-context matters more than the peak number. For sizing that model's memory footprint properly, see our Qwen3.8-27B hardware guide — the KV cache at this model's native context is larger than its weights.

A caution on the other numbers circulating this month: the widely-shared NordicSilicon Mac mini benchmark table (7 September 2026) is explicitly not measurement. Its own methodology note reads "M6, M5 Pro, M5 Max and M5 Ultra figures are projections scaled from published memory bandwidth and prefill characteristics — retail units ship September 22, 2026." Projections scaled from bandwidth are a reasonable modelling exercise and a terrible basis for a $1,000 purchase decision. Do not treat them as benchmarks.

Two supported models is the real catch

Splash supports exactly two models: incoai/Qwen3.8-27B-Splash and Qwen3.6-35B-A3B. That's the list.

This is the design, not a launch limitation — Splash is built around specific model architectures rather than being a general runtime that happens to load GGUFs. That is where the 2–3x comes from. It is also the thing that should govern the purchase, and it is the thing vendor posts structurally bury.

Concretely: if you configure a 48GB Mac mini specifically to run Splash, you are spending the memory upgrade to accelerate two models. Everything else in your stack — Llama, Gemma, DeepSeek, image models, embedding models, whatever you run next quarter — routes through MLX or llama.cpp as before, at their normal speeds. Splash is an accelerator for one lane, not a replacement runtime.

That is not a reason to skip it. It is a reason to buy the 48GB machine because 48GB is the right amount of memory for local AI generally, and treat Splash as a strong bonus on top. Buy for capability, not for one engine's supported-model list. Our how much RAM do you need for local AI guide makes that case independently of Splash.

36GB of unified memory vs 24GB of VRAM at the same money

Splash is good enough to genuinely pull cross-shoppers toward Apple. The 36GB tax is what pushes some of them back. Both branches are defensible, so here is the decision rule rather than a winner.

The NVIDIA side of this got faster this month too, which matters for the honest comparison. At IFA 2026 NVIDIA announced kernel-level optimizations claiming "llama.cpp delivers up to 1.9x higher throughput through kernel optimizations on a GeForce RTX 5090", plus 1.2x on vLLM for the RTX PRO 6000 Blackwell Workstation Edition. Vendor-reported, same caveat as Inco's — but it means the "Apple just got 2–3x faster" framing is not happening in a vacuum. See our 2026 GPU price guide for where these cards actually sell.

The Strix Halo branch deserves its own note. A 128GB EVO-X2 has more than twice the memory of a maxed M5 Pro mini, at a price between the mini and the Studio — but its unified LPDDR5X bandwidth trails both Apple and a discrete card, and the ROCm path is less mature than CUDA or Metal. Buy it when capacity is the constraint and you are patient. Details in our Strix Halo mini PC guide and the broader mini PC for AI hub.

One route this post deliberately does not recommend for Splash: clustering. If you are at 32GB and tempted to add a second box rather than buy more memory, read Mac mini clustering for local AI first — but note that Splash's floor is a per-machine requirement. Two 32GB minis are two ineligible machines, not one 64GB eligible one.

Eligibility checklist and what to buy

That last row is the one we would want a friend to read. A memory upgrade is the single most expensive line item on an Apple configurator and it is irreversible. "An engine I can't run supports two models I may not use" is a weak reason to pay it. "48GB is the right amount of unified memory for local AI in 2026 and Splash is a bonus" is a strong one. If you own an M4 Pro mini at 24GB, you clear the chip gate and miss the memory gate — which is annoying, but does not make your machine worse at everything else it was doing yesterday.

For the wider Apple ladder, our M5 Mac mini and Mac Studio guide covers capacity versus prompt-processing tradeoffs across the full lineup, and the Apple Silicon for AI hub collects everything. If Splash pushes you off Apple entirely, Mac mini alternatives for AI and best mini PC for AI cover the other side.

How to check your own Mac in 30 seconds

Three conditions, one screen. Click the Apple menu → About This Mac:

1. Chip — must read M3, M4, M5, M6, or any Pro/Max/Ultra variant of those. M1 and M2 fail regardless of memory.
2. Memory — must read 36 GB or higher. 32GB fails. There is no rounding here.
3. macOS — must be 26.4 or later. If it isn't, System Settings → General → Software Update fixes it for free.

All three pass? Install Homebrew, open LM Studio, and Splash is available to you at no cost. Any one fails? The chip and memory conditions are purchase decisions, not settings — and the table above tells you which purchase.

For a deeper dive on whether a given model fits a given machine — the floor is one thing, actual allocatable memory after macOS takes its share is another — run your target model through our memory-fit tool and read the local LLM guide hub.

Bottom line

Splash is a real advance and a narrow one. It is 2–3x faster than the alternatives on two models, on a subset of Macs, behind a memory floor that the month's most popular new Mac cannot reach at any configuration.

If you own a 36GB+ M3-or-newer Mac: install it today, it costs nothing. If you're buying a mini: the M5 Pro at 48GB, which is $2,299 all-in once the $600 memory upgrade is counted — and at that number, price the $2,499 Mac Studio M5 Max against it before you commit. If you're at 32GB or below: this is not the reason to spend, and anyone telling you the $899 M6 mini gets you here is wrong.