LLMBOTTLENECK.COM

Can the AMD Instinct MI355X run Kimi-K3?

Kimi-K3 · AMD Instinct MI355X · 8,192 ctx · whole model residentQ2_K sizesize range

IT DOES NOT FIT

No. Even Q2_K, the smallest format evaluated, needs 1110.3 GB against 288 GB — short by 822.3 GB. No published file exists for this pair, so every size is rebuilt from the pinned architecture.

0/9formats that fit
288 GBdevice memory
1110.3 GBneeded at smallest format
—speed unavailable: does not fit
8000GB/s bandwidth
Run it in the cloud

Run Kimi-K3 anyway, on a rented GPU

The AMD Instinct MI355X cannot hold it, but these rented machines hold the whole model. The ones worth choosing between, priced live and billed by the second.

Your card Offloading would put about 1442 GB in system RAM — more than 64 GB, and at system-memory speed. For this model, a machine that holds it all on the GPU is the practical way to run it. Check with your RAM →

Prices · 18:30 UTC, 22 Sept

The machines that hold Kimi-K3 at Q4_K_M are not in stock at Vast.ai or RunPod right now.

No GPU at all

Use Kimi-K3 by the token

Nothing to set up and nothing to switch off: you pay only for the tokens you use.

Every machine holds the whole model at Q4_K_M and 8,192 tokens of context, no offload. Speeds are this site’s single-stream decode estimates; prices are what each provider’s own API quoted, on-demand, for the whole machine. Vast hosts below 98% measured reliability are left out.

Referral links Vast.ai, RunPod and Novita pay us a share of what you spend if you sign up through these buttons. It costs you nothing, and it never decides an order or a recommendation: both are computed from the live price and the speed, and options that pay us nothing are listed and recommended on the same terms. How we rank

Runs well on your AMD Instinct MI355X instead

The models nearest to Kimi-K3 — same lab first, then closest in size — that the AMD Instinct MI355X holds whole at the standard format and runs at a usable speed. Nearest in size is not the same as equally good; compare them on the task you care about.

Every format evaluated

No published file exists for this pair, so every size is rebuilt from the pinned architecture. The Basis column says which is which for each row. Decode and first-token figures are calibrated estimates, not runs on this card.

FormatNeedsOf 288 GBFitsDecodeFirst tokenBasis
FP165604.0 GB1946%short by 5316.0 GBdoes not run—size range
Q8_02979.0 GB1034%short by 2691.0 GBdoes not run—size range
Q6_K2301.0 GB799%short by 2013.0 GBdoes not run—size range
Q5_K_M2001.8 GB695%short by 1713.8 GBdoes not run—size range
Q5_01955.6 GB679%short by 1667.6 GBdoes not run—size range
Q4_K_M1720.4 GB597%short by 1432.4 GBdoes not run—size range
Q4_01630.8 GB566%short by 1342.8 GBdoes not run—size range
Q3_K_M1401.9 GB487%short by 1113.9 GBdoes not run—size range
Q2_K1110.3 GB386%short by 822.3 GBdoes not run—size range

Sized at 8,192 tokens of context with the whole model resident — weights, the KV cache and the runtime reserve, offload off. Speed is only quoted for a format that fits: a rate for a configuration that cannot load is not a fact about anything.

Why the first-token figure is the same on every row

First token is modelled from the arithmetic the prompt requires, and that count does not change with the weight format — which is why it reads the same on every row. Real prefill does vary by format, because a quantized matmul is a different kernel; this model does not capture that, and the figure should be read as an order of magnitude rather than a ranking between formats.

What the labels mean
published data
Read from a published source — a file's byte count, a model's configuration or a manufacturer's specification — or exact arithmetic on such values. Not a measurement on a machine.
reconstructed size
Weight size reconstructed from the pinned architecture, because no published file exists.
size range
Only a lower and an upper bound are claimed for this weight size.
estimate
Calculated from sourced inputs by a stated method; an estimate, not a measurement.

Ways out

OPTION 1cluster

4× AMD Instinct MI355X

4 of the same card, tensor parallel, holds it at Q2_K. Often cheaper than one larger card, and it is the option most sizing tools never mention.

Open that configuration →
OPTION 2system RAM

Offload to system memory

Keeping some layers on the host lets it load at all, at a large cost in speed. The calculator sizes the split and prices it.

Size the offload →
What this device does run

The full list is on the AMD Instinct MI355X page.

Change anything

This page fixes the context at 8,192 tokens and one device. Batch, concurrent users, KV-cache format, clusters and rental cost are all in the calculator, already set to this pairing.

Open in the calculator →Everything the AMD Instinct MI355X runs

Where these numbers come from

Kimi-K3’s architecture is read from its publisher’s own config.json at a pinned revision, and the AMD Instinct MI355X’s 288 GB and 8000 GB/s come from the manufacturer’s specification. How far each figure can be trusted is published, per format and worst case included, on the accuracy scorecard.

Building this into your own product? A free API key returns exactly these figures, and the widget puts this answer on a product page with one script tag.

Use this answer in your own product

As JSON, from the API
curl -s https://llmbottleneck.com/v1/analyze \
  -H "Authorization: Bearer $LLMB_KEY" -H "content-type: application/json" \
  -d '{"model":"moonshotai-kimi-k3","quantization":"Q2_K","context":8192,"hardware":"amd-instinct-mi355x"}'

Same engine, same evidence, every field sourced. Free key in one step →

On your page, as a widget
<script src="https://llmbottleneck.com/widget.js"
  data-model="moonshotai-kimi-k3" data-quantization="Q2_K"
  data-hardware="amd-instinct-mi355x" data-context="8192"></script>

No key needed. Unbranded, with your own buy button, on Pro and Business →

llmbottleneck
catalogue 2026-10-03models 327devices 135