NVMe speed and local LLM model loading times

Running a language model locally is now a normal workflow for developers, researchers, small studios and anyone who does not want prompts leaving the building. The catch is model size. A quantised Llama 3 70B lands around 40GB, Mixtral 8x7B in Q5 sits close to 32GB, and even a 34B coder model is 20GB before you count context cache. Every one of those files lives on the SSD until inference starts, and that is when the drive matters.

The one number that decides loading speed

For LLM loading, the spec that decides the wait is sequential read. Random 4K IOPS, controller cache and DRAM configuration all matter for general Windows responsiveness, but a model file is one massive blob read end to end into system RAM or VRAM. It is close to a best case scenario for a modern NVMe controller, which means the drive will hit rated sequential read speed for almost the entire duration of the load.

Rated sequential read numbers for the mainstream tiers in 2026 look like this. Gen 3 NVMe tops out around 3,500 MB/s. Gen 4 sits in the 5,000 to 7,400 MB/s band depending on model. Gen 5 pushes into 10,000 to 14,500 MB/s territory on drives like the Crucial T705, Samsung 9100 Pro and Corsair MP700 Pro.

What that means in seconds, not megabytes

Take a 30GB model file, which is roughly a 34B parameter model at Q6 or a 70B model at Q3. Divide file size by real world sustained read speed and you get a defensible loading floor.

On a Gen 3 drive at 3,500 MB/s you are looking at about 8.5 seconds of pure disk read. On a Gen 4 drive at 7,000 MB/s that drops to just under 4.4 seconds. On a Gen 5 drive holding 12,000 MB/s the same file loads in roughly 2.5 seconds.

Push the file up to 45GB, roughly a 70B Q4 model, and the gap widens. Gen 3 needs about 13 seconds. Gen 4 needs around 6.5 seconds. Gen 5 finishes in under 4. None of those numbers include the CPU and RAM overhead of unpacking the format, mapping tensors and moving weights to the GPU, all of which adds a second or two at the front end regardless of drive.

Why real world numbers are slower than the box

Rated speeds assume a fresh drive, an empty SLC cache and a PCIe slot wired directly to the CPU. In practice most builds see 80 to 90 percent of the rated figure once the drive is more than half full. Gen 5 drives in particular throttle hard without a proper heatsink, and a throttled Gen 5 can end up slower than a well cooled Gen 4 during a long sustained read.

The other quiet killer is chipset lanes. If your motherboard routes the second M.2 slot through the chipset, a Gen 4 drive in that slot behaves like a Gen 3. Always put the primary model drive in the top M.2 slot.

When Gen 5 is actually worth the money

If you load one model at the start of the day and leave it resident, the difference between Gen 3 and Gen 5 is a rounding error across a working week. If you swap between three or four models an hour, testing prompts against different quantisations or routing tasks to different specialist models, those seconds add up into real minutes of interrupted flow.

Gen 5 also earns its price when tools stream shards from disk on demand rather than loading the full model up front. For a straightforward llama.cpp or Ollama workflow with a 30GB model that sits in RAM, a good Gen 4 drive is hard to beat on cost per useful second.

See the BAV 2TB NVMe build on eBay

The BAV 2TB NVMe recommendation

The Birmingham AV 2TB NVMe build ships with a Gen 4 drive that hits sustained sequential reads in the 6,800 to 7,200 MB/s range on the top M.2 slot. That puts a 30GB model load at roughly 4.3 seconds of disk time, and a 45GB model load at just under 7. It is the tier that turns local LLM work from something you plan around into something you barely notice.

Two terabytes is also the point where model storage stops being a constant tidying job. A working developer can keep four or five model families resident, plus quantised variants, plus Windows and a decent games library, without the constant shuffle that 1TB forces.

Depending on the configuration selected at checkout, the build covers a range of CPU, RAM and GPU pairings, and the 2TB Gen 4 NVMe is the storage baseline across that range. If the variation you pick specifies a Gen 3 drive rather than Gen 4, expect the loading numbers above to roughly double.

FAQ

Does more system RAM make the drive matter less?

Only for the second load onwards. The OS caches the model file in unused RAM after the first read, so repeat loads come off memory regardless of drive speed. The first load of the day, and every load after a reboot or a model swap, still hits the SSD at full sequential read.

Will a Gen 5 drive help GPU inference speed?

No. Once weights are loaded into VRAM the SSD is idle. Inference speed is decided by GPU compute, memory bandwidth and quantisation. The drive only matters at the moment of loading, and any time the model is partially offloaded to system RAM and paged.

Is 1TB enough for local LLM work?

For one model family at a time, yes. For anyone experimenting with multiple architectures, keeping a few quantisation levels of each, or working with image and audio models alongside text, 1TB fills up faster than expected. 2TB is the practical minimum for serious local work in 2026.

Does a heatsink actually matter for LLM loading?

For short 30GB loads, not much. The read completes before the drive hits its throttle threshold. For 45GB and larger models, or for workflows that stream weights continuously, a proper heatsink keeps the drive at rated speed for the full duration and prevents the sudden slowdowns that show up in longer sessions.

About Birmingham AV

Birmingham AV has sold more than 87,000 items on eBay since 2017, with 24,756 buyer feedbacks at 98.9 percent positive. Every system ships with a twelve month warranty. Companies House number 12383651, VAT registration GB 348755066, based in Bromsgrove, Worcestershire. One of the highest volume refurbished PC operations on eBay UK, and every build is tested, cleaned and configured in house before it leaves the workshop.