LLMBOTTLENECK.COM
Meta / llama / file from the unsloth mirror

Llama-3.2-3B-Instruct

3.21 billion parameters.

Open in the calculator →

Architecturepublished data
Quick answer
Llama-3.2-3B-Instruct needs about 3.98 GB at Q4_K_M — the format most people download — with 8,192 tokens of context. It fits entirely on 18 of 18 common devices, starting with the NVIDIA GeForce RTX 5070 (12 GB). What else fits in 12 GB. No hardware for it? Rent a GPU that runs it, priced live.

Sized at 8,192 tokens of context with the whole model in device memory, across 18 common devices. Speeds are estimates, not benchmarks.

Find the NVIDIA GeForce RTX 5070, the smallest card listed here that holds it: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

Configuration read from a mirror of a restricted repository

The official repository is meta-llama/Llama-3.2-3B-Instruct (licence: llama3.2). Meta's repository is licence-gated. This configuration is read from Unsloth's ungated copy, which declares the official repository as its base model; it is not byte-compared with Meta's file.

Weight-file size by format

FormatSizeRangeBasis
FP167.22 GB6.43 GB – 7.26 GBreconstructed size
Q8_03.84 GB3.41 GB – 3.86 GBreconstructed size
Q6_K2.96 GB2.64 GB – 2.98 GBreconstructed size
Q5_K_M2.59 GB2.26 GB – 2.64 GBreconstructed size
Q5_02.54 GB2.21 GB – 2.59 GBreconstructed size
Q4_K_M2.24 GB1.91 GB – 2.34 GBreconstructed size
Q4_02.14 GB1.81 GB – 2.24 GBreconstructed size
Q3_K_M1.86 GB1.54 GB – 2.02 GBreconstructed size
Q2_K1.46 GB1.14 GB – 1.66 GBreconstructed size

A file size is not the memory a run needs: the KV cache and the runtime reserve come on top, and the calculator adds both.

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.

Hardware ladder

Q4_K_M at 8,192 tokens · 18 of 18 devices hold it entirely
DeviceMemoryVerdictDecodeCalculator
NVIDIA GeForce RTX 50703.98 GB of 12 GBfits~155 tok/sOpen →
AMD Radeon™ RX 9070 XT3.98 GB of 16 GBfits~169 tok/sOpen →
NVIDIA GeForce RTX 40803.98 GB of 16 GBfits~165 tok/sOpen →
NVIDIA GeForce RTX 50803.98 GB of 16 GBfits~221 tok/sOpen →
NVIDIA GeForce RTX 5070 Ti3.98 GB of 16 GBfits~206 tok/sOpen →
NVIDIA GeForce RTX 5060 Ti 16GB3.98 GB of 16 GBfits~103 tok/sOpen →
NVIDIA GeForce RTX 40903.98 GB of 24 GBfits~232 tok/sOpen →
AMD Radeon™ RX 7900 XTX3.98 GB of 24 GBfits~253 tok/sOpen →
NVIDIA GeForce RTX 30903.98 GB of 24 GBfits~215 tok/sOpen →
NVIDIA GeForce RTX 50903.98 GB of 32 GBfits~412 tok/sOpen →
NVIDIA RTX 6000 Ada Generation3.98 GB of 48 GBfits~221 tok/sOpen →
Apple M4 Pro3.98 GB of 64 GBfits~58 tok/sOpen →
Apple M5 Pro3.98 GB of 64 GBfits~63 tok/sOpen →
AMD Ryzen AI Max+ 395 with Radeon 8060S3.98 GB of 128 GBfits~68 tok/sOpen →
Apple M3 Max3.98 GB of 128 GBfits~74 tok/sOpen →
Apple M4 Max3.98 GB of 128 GBfits~87 tok/sOpen →
NVIDIA DGX Spark3.98 GB of 128 GBfits~63 tok/sOpen →
Apple M2 Ultra3.98 GB of 192 GBfits~104 tok/sOpen →

Decode is a calibrated estimate from the published calibration; capacity is the manufacturer's published ceiling, not guaranteed free memory. Fitting in memory is not the same as loading: whether the runtime and version you have supports this architecture and format on that machine has not been tested here. Offloaded rows assume there is enough system RAM for the overflow — the calculator checks that against the RAM you declare. Each “Open” link keeps this model, format and context.

Run it in the cloud

Rent a GPU that runs Llama-3.2-3B-Instruct

Every rentable machine that holds the whole model, from Vast.ai and RunPod, with this site's speed estimate for each and the provider's own price, read live. Cheapest first; sort by cost per token or speed instead, or switch the format.

Prices · 18:30 UTC, 22 Sept
28 configurations in stock · 5 out of stock
MachineSpeedPer hourPer M tokensWhereRent
1× RTX 306012 GBrecommendedcheapestbest value~82 tok/sFaster than you read$0.036$0.12Vast.aiMarketplace · 99.8% reliableRent on Vast.ai
1× RTX 308010 GB~174 tok/sFaster than you read$0.082$0.13Vast.aiMarketplace · 98.3% reliableRent on Vast.ai
1× RTX 30708 GB~103 tok/sFaster than you read$0.082$0.22Vast.aiMarketplace · 99.1% reliableRunPod $0.13/h Rent on Vast.ai
1× RTX 4060 Ti 16GB16 GB~66 tok/sFaster than you read$0.090$0.38Vast.aiMarketplace · 99.4% reliableRent on Vast.ai
1× RTX 407012 GB~116 tok/sFaster than you read$0.096$0.23Vast.aiMarketplace · 99.8% reliableRent on Vast.ai
1× RTX 309024 GB~215 tok/sFaster than you read$0.12$0.16Vast.aiMarketplace · 99.5% reliableRunPod $0.50/h Rent on Vast.ai
1× RTX 5060 Ti 16GB16 GB~103 tok/sFaster than you read$0.12$0.33Vast.aiMarketplace · 99.1% reliableRent on Vast.ai
1× RTX 3080 Ti12 GB~209 tok/sFaster than you read$0.14$0.18Vast.aiMarketplace · 99.9% reliableRent on Vast.ai

No GPU at all

Use Llama-3.2-3B-Instruct by the token

Here the rented card is as cheap as the API or cheaper per token ($0.12 per million), and it keeps your data on a machine you control. The API is still the easier start: nothing to set up and nothing to switch off.

First time renting a GPU? How it works, in four steps
  1. Create an account and add credit. Both providers are prepaid: $10 is enough to try any card on this page for hours. A new RunPod account that arrives through a referral link gets a one-time $5 credit when it first adds $10.
  2. Pick the card and the count shown here, and a template: llama.cpp or vLLM for an OpenAI-compatible API, or one with a web chat if you only want to talk to the model. “How to launch it” below gives the exact image and command.
  3. Wait for the download. The weights are fetched on the machine; tens of gigabytes take a few minutes on a data-centre connection. You pay by the second from the moment it starts.
  4. Stop it when you are done. A running machine bills even when idle. A stopped one costs nothing per hour, though its disk is usually still billed until you delete it.
How to launch it

No Q4_K_M file of Llama-3.2-3B-Instruct is catalogued here, so there is no exact command to vouch for. Start the machine from the llama.cpp server image ghcr.io/ggml-org/llama.cpp:server-cuda and point -hf at a Q4_K_M build from the publisher or a quantizer you trust — search Hugging Face for one. Check its file size against the memory figure on this page before you rent.

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

Buying a card instead, or paying by the token? Compare a year of Llama-3.2-3B-Instruct three ways →

Under the hood

Architecture, read from the publisher’s file

The numbers every figure above is computed from, with the file they came from.

Architecture

✓ Architecture read from the mirror's config.json

Retrieved 2026-09-13 at pinned commit 006f5dcd1393; parameter count from the safetensors index at the same revision.

Architecture
llama
Layers
28
Hidden size
3,072
Attention heads
24
KV heads
8
Head dimension
128
Feed-forward width
8,192
Vocabulary
128,256
Context ceiling
131,072
RoPE theta
500,000

Where the memory goes

tokenembedding× 28 decoder blocksGrouped-query attention24 query · 8 KV headsfull contextFeed-forwardone networkall activeoutputprojectiongrows with contextfixed per token

Grouped-query attention shares each key/value head across 3 query heads, so the KV cache is 33% of what multi-head attention would need at the same context.

Open the exact config.json this page was read from (the mirror) ↗

More from Meta

Llama-3.2-1B-InstructLlama-3.3-70B-InstructLlama-4-Maverick-17B-128E-InstructLlama-4-Scout-17B-16E-InstructMeta-Llama-3.1-70B-Instruct

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":"unsloth-llama-3-2-3b-instruct","quantization":"Q4_K_M","context":8192,"hardware":"nvidia-geforce-rtx-4090"}'

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="unsloth-llama-3-2-3b-instruct" data-quantization="Q4_K_M"
  data-hardware="nvidia-geforce-rtx-4090" data-context="8192"></script>

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

Citing this page

LLM Bottleneck. “Llama-3.2-3B-Instruct VRAM and hardware requirements.” Architecture from unsloth/Llama-3.2-3B-Instruct at revision 006f5dcd1393, retrieved 2026-09-13. https://llmbottleneck.com/models/unsloth-llama-3-2-3b-instruct

Every figure above is either the published value or a reconstruction whose measured error is on the accuracy page.

llmbottleneck
catalogue 2026-10-03models 327devices 135