InferSSD on a 256 GB Mac Studio, Tested: Plain llama.cpp Wrote Three Times Faster

By Evan Vega

WiCi’s open-source InferSSD engine, built to run AI models far larger than a computer’s memory, wrote answers at a third
of the speed of the standard llama.cpp engine on a Mac Studio with 256 GB of memory, Novel Cognition found in an overnight
test on 1-2 November 2026. Both engines ran the same DeepSeek V4.1 Flash model, about 748 billion parameters and a 473 GiB
file, using InferSSD’s own benchmark script. llama.cpp wrote at 10.95 tokens a second; InferSSD’s best setting wrote at
3.45. InferSSD read the prompt as fast as llama.cpp or faster.

The test follows Novel Cognition’s check of NetworkChuck’s WiCi One video, which was sponsored by WiCi. The full results,
with charts, a red team of the result and the method, are at inferssd.novcog.us.com.

Novel Cognition ran WiCi’s open-source InferSSD engine and the standard llama.cpp engine on the same 473 GiB DeepSeek V4.1 Flash model on a Mac Studio with 256 GB of memory. llama.cpp wrote answers at

What was checked, and what was not

Both engines ran the same file, used the same six experts per token, built the model’s lookup tables into every token,
ran without speculative decoding and received identical input token ids. Novel Cognition’s own pre-registered prediction,
that InferSSD would be at least twice as fast as llama.cpp on the Studio, was wrong. WiCi’s published Mac result, a 48 GB
MacBook Pro on which InferSSD wrote 3.29 times faster than the same llama.cpp fork, is a different machine and is not
contradicted: a 48 GB laptop has no room to cache the model.

Not yet tested: repeat runs of llama.cpp, runs with the cache cleared between them, either engine sharing the machine
with another loaded model, and the model on internal rather than Thunderbolt storage. The red team
and the scored predictions are on the analysis site.

Writing speed, by memory size

InferSSD keeps its own cache of the model’s most-used expert blocks inside a memory budget the user sets. Its writing
speed rose with that budget: 1.96 tokens a second at 16 GiB, 2.20 at 32 GiB, 2.74 at 56 GiB and 3.45 at 112 GiB. At a
200 GiB budget the run was stopped all three times by the benchmark’s own memory guard once macOS began compressing the
engine’s cache; the launcher’s 64 GiB reserve meant that setting allocated about 157 GiB.

llama.cpp, which simply maps the model file and lets macOS cache what it reads, ended its run with about 223 GiB of the
model in the operating system’s file cache, read 0.023 GB from the drive per token written, and caused no memory
compression or swapping. InferSSD at 112 GiB read 0.37 GB per token.
Why the Mac’s own cache won and why the largest setting never ran.

InferSSD is the faster reader

Before a model writes, it reads the prompt. On the 4,104-token test prompt InferSSD read at 12.3 to 14.7 tokens a
second across its settings, against llama.cpp’s 12.5, while streaming 770 to 880 GB from the drive at close to the
drive’s full speed; llama.cpp read about 133 GB from the drive. In one run that started, by accident, with the Mac’s cache
already warm, InferSSD read at 31.1 tokens a second, with 32 reading threads against llama.cpp’s 24.

Which engine finishes a request first therefore depends on answer length. On the measured speeds, llama.cpp pulls ahead
of InferSSD’s 112 GiB setting after roughly 130 to 150 answer tokens. DeepSeek V4.1 Flash reasons before answering, so its
answers usually run far longer than that. The two speeds, charted.

What it means

Novel Cognition’s reading: on a small machine that cannot hold the model, WiCi’s published numbers favour InferSSD; on a
big-memory Mac given the whole machine, the standard engine writes about three times faster; for everyday use, a smaller
model that fits in memory is faster than both. The method is at method, and the test is also an
episode of the Frontier Watch podcast.

Novel Cognition runs this kind of vendor-claim testing for companies as part-time AI leadership: which models to run,
on what hardware, and what to build or buy (novcog.com/contact). The checks behind
every published number are at hiddenstatedrift.com/verify. Novel
Cognition is not affiliated with WiCi, NetworkChuck, DeepSeek or Apple, and nobody paid for this test.

Part of the Frontier Watch Series: Read the previous investigation

More Coverage:
→ Read this investigation on North Denver Tribune
→ Coverage from Daily California Press