IBMGraniteGraniteForCausalLM

Granite 3.3 2B Instruct — Hardware Requirements & GPU Compatibility

ChatFunctions

Granite 3.3 2B Instruct is a 2.5B-parameter open language model from IBM in the Granite family. It supports a context window of up to 131,072 tokens. At Q4_K_M it needs about 1.99 GB of VRAM — see which GPUs and Macs can run it below.

2.5B params · dense131K contextApache 2.0Apr 202578.6K downloads87 likes11.1K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Runs great on a 24 GB GPU

Runs from 8 GB (Q4_K_M · 64K)Best on 24 GB (Q4_K_M · ~329 tok/s · 128K)Macs from 8 GB (Q4_K_M · ~36 tok/s)

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Granite 3.3 2B Instruct

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Ways to runOther ways

each opens step-by-step

Run in cloud

Runs on RTX 3080 Ti (12 GB) (8 GB headroom) · Q4_K_M

Generation speed: ~298 tok/s; generation speed
Cost per 1M output tokens: $0.11; per 1M output tokens

Deployon Vast.aiBest valuefrom $0.12/hr

Compare GPUs →

Run Granite 3.3 2B Instruct

3.2 GB VRAM at 16K context · fits the RTX 4090 (20.8 GB free)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/ibm-granite/granite-3.3-2b-instruct-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf ibm-granite/granite-3.3-2b-instruct-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/ibm-granite/granite-3.3-2b-instruct-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf ibm-granite/granite-3.3-2b-instruct-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve ibm-granite/granite-3.3-2b-instruct --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~6 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model mlx-community/granite-3.3-2b-instruct-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model mlx-community/granite-3.3-2b-instruct-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$ibm-granite/granite-3.3-2b-instruct-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
RTX 3080 Ti (12 GB)CheapestCheapest8 GB free	~298 tok/s	$0.11	Deploy on Vast.aifrom $0.12/hr (opens in new tab)	12 GB · 8 GB free · ~298 tok/s · $0.11 / 1M tokens
RTX 3060 12GB8 GB free	~118 tok/s	$0.15	Deploy on Vast.aifrom $0.061/hr (opens in new tab)	8 GB free · ~118 tok/s · $0.15 / 1M tokens
RTX 4060 (8 GB)4 GB free	~89 tok/s	$0.27	Deploy on Vast.aifrom $0.086/hr (opens in new tab)	8 GB · 4 GB free · ~89 tok/s · $0.27 / 1M tokens
RTX A4000 (16 GB)12 GB free	~146 tok/s	$0.17	Deploy on Vast.aifrom $0.088/hr (opens in new tab)	16 GB · 12 GB free · ~146 tok/s · $0.17 / 1M tokens
RTX 4070 SUPER (12 GB)8 GB free	~165 tok/s	$0.17	Deploy on Vast.aifrom $0.10/hr (opens in new tab)	12 GB · 8 GB free · ~165 tok/s · $0.17 / 1M tokens
RTX PRO 4500 (32 GB)28 GB free	~293 tok/s	$0.32	Deploy on RunPodfrom $0.34/hr (opens in new tab)	32 GB · 28 GB free · ~293 tok/s · $0.32 / 1M tokens
RTX 4090 (24 GB)20 GB free	~329 tok/s	$0.29	Deploy on RunPodfrom $0.34/hr (opens in new tab)	24 GB · 20 GB free · ~329 tok/s · $0.29 / 1M tokens
A40 (48 GB)44 GB free	~227 tok/s	$0.43	Deploy on RunPodfrom $0.35/hr (opens in new tab)	48 GB · 44 GB free · ~227 tok/s · $0.43 / 1M tokens
L4 (24 GB)20 GB free	~98 tok/s	$1.25	Deploy on RunPodfrom $0.44/hr (opens in new tab)	24 GB · 20 GB free · ~98 tok/s · $1.25 / 1M tokens

Compare GPUs →

How Much VRAM Does Granite 3.3 2B Instruct Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
1.2 GB	3.7 GB	11.7 GB	0.7 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 128K context · ~565 tok/s
1.2 GB	3.8 GB	11.8 GB	0.8 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 128K context · ~533 tok/s
1.3 GB	3.8 GB	11.8 GB	0.8 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 128K context · ~520 tok/s
1.3 GB	3.8 GB	11.9 GB	0.9 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 128K context · ~496 tok/s
1.4 GB	4 GB	12 GB	1 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 128K context · ~452 tok/s
1.5 GB	4 GB	12.1 GB	0.9 GB	2-bit · small K-quant — 2-bit small K-quant	Fits: 128K context · ~443 tok/s
1.5 GB	4 GB	12.1 GB	1.1 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 128K context · ~434 tok/s
1.5 GB	4.1 GB	12.1 GB	1 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 128K context · ~426 tok/s
1.5 GB	4.1 GB	12.1 GB	1.1 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 128K context · ~426 tok/s
1.6 GB	4.1 GB	12.2 GB	1.1 GB	3-bit small — 3-bit small quantization	Fits: 128K context · ~415 tok/s
1.6 GB	4.1 GB	12.2 GB	1.2 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 128K context · ~407 tok/s
1.7 GB	4.2 GB	12.3 GB	1.3 GB	3-bit medium — 3-bit medium quantization	Fits: 128K context · ~385 tok/s
1.7 GB	4.3 GB	12.3 GB	1.5 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 128K context · ~379 tok/s
1.8 GB	4.3 GB	12.3 GB	1.4 GB	3-bit large — 3-bit large quantization	Fits: 128K context · ~370 tok/s
1.8 GB	4.4 GB	12.4 GB	1.4 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 128K context · ~358 tok/s
1.9 GB	4.4 GB	12.5 GB	1.5 GB	4-bit small — 4-bit small quantization	Fits: 128K context · ~347 tok/s
1.9 GB	4.4 GB	12.5 GB	1.6 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Fits: 128K context · ~347 tok/s
1.9 GB	4.4 GB	12.5 GB	1.5 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Fits: 128K context · ~347 tok/s
2.0 GB	4.5 GB	12.6 GB	1.5 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 128K context · ~329 tok/s
2.0 GB	4.5 GB	12.6 GB	1.6 GB	4-bit large — 4-bit large quantization	Fits: 128K context · ~324 tok/s
2.0 GB	4.6 GB	12.6 GB	1.8 GB	5-bit · legacy — 5-bit legacy quantization	Fits: 128K context · ~320 tok/s
2.2 GB	4.7 GB	12.8 GB	1.8 GB	5-bit small — 5-bit small quantization	Fits: 128K context · ~297 tok/s
2.2 GB	4.7 GB	12.8 GB	1.9 GB	5-bit · legacy+ — 5-bit legacy quantization with offset	Fits: 128K context · ~297 tok/s
2.3 GB	4.8 GB	12.8 GB	1.8 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Fits: 128K context · ~289 tok/s
2.3 GB	4.8 GB	12.9 GB	1.8 GB	5-bit large — 5-bit large quantization	Fits: 128K context · ~285 tok/s
2.6 GB	5.1 GB	13.1 GB	2.1 GB	6-bit, very good — 6-bit quantization, very good quality	Fits: 128K context · ~256 tok/s
3 GB	5.5 GB	13.6 GB	2.7 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Fits: 128K context · ~218 tok/s
5.5 GB	8.1 GB	16.1 GB	5.1 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Fits: 128K context · ~119 tok/s

Selected Q4_K_M

4-bit medium quantization

3.2GBat 16K context

Weights: 1.5 GB
KV cache: 1.3 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · 20.8 GB free · ~329 tok/s

≥50 tok/s needs ≥153 GB/s · RTX 5090 ~585, RTX 3090 Ti ~329

Which Hardware Can Run Granite 3.3 2B Instruct?

Q4_K_M · 2.0 GB

Granite 3.3 2B Instruct (Q4_K_M) requires 2.0 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 3+ GB is recommended. Using the full 131K context window can add up to 10.6 GB, bringing total usage to 12.6 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 163 devices with unified memory can run Granite 3.3 2B Instruct, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8 GB · RTX 4060Q4_K_M · 3.2 GB · ~89 tok/s · 64K3.2~8964K8 GB tier →

12 GB · RTX 3060 12GBQ4_K_M · 3.2 GB · ~118 tok/s · 64K3.2~11864K12 GB tier →

16 GB · RTX 5060 Ti 16GBQ4_K_M · 3.2 GB · ~146 tok/s · 128K3.2~146128K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 3.2 GB · ~329 tok/s · 128K3.2~329128K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 3.2 GB · ~42 tok/s · 128K3.2~42128KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 3.2 GB · ~192 tok/s · 128K3.2~192128KMacs →

Alternatives to Granite 3.3 2B Instruct

Granite family · 28 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
WeVisOne VL 2BTencent	—	3.6 GB	~328	same size
MiniCPM5 2BOpenBMB	—	2.5 GB	~345	same size
North Micro Vision InstructCohere	—	3.7 GB	~323	same size
LFM2.5 2.6BLiquid AI	—	2.3 GB	~321	same size
Gemma 2 2B ITGoogle	—	2.5 GB	~379	same size
Ouro 2.6B ThinkingByteDance	—	12.1 GB	~102	same size

Specifications

Publisher: IBM
Family: Granite
Parameters: 2.5B
Architecture: GraniteForCausalLM
Context Length: 131,072 tokens
Vocabulary Size: 49,159
Release Date: 2025-04-09
License: Apache 2.0

Resources

Frequently Asked Questions

How much VRAM does Granite 3.3 2B Instruct need?

Granite 3.3 2B Instruct requires 2.0 GB of VRAM at Q4_K_M, or 5.5 GB at BF16. Full 131K context adds up to 10.6 GB (12.6 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 2.5B × 4.8 bits ÷ 8 = 1.5 GB

KV Cache + Overhead ≈ 0.5 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 11.1 GB (at full 131K context)

VRAM usage by quantization

Q4_K_M

2.0 GB

Q4_K_M + full context

12.6 GB

Learn more about VRAM estimation →

What's the best quantization for Granite 3.3 2B Instruct?

For Granite 3.3 2B Instruct, Q4_K_M (2.0 GB) offers the best balance of quality and VRAM usage. Q4_K_L (2.0 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 1.2 GB.

VRAM requirement by quantization

IQ2_XXS

1.2 GB~53%

Q2_K

1.5 GB~75%

IQ4_XS

1.8 GB~87%

Q4_K_M ★

2.0 GB~89%

Q5_0

2.0 GB~90%

BF16

5.5 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Granite 3.3 2B Instruct on a Mac?

Yes — MacBook Air 13" M3 (8 GB) and 145 other Macs can run Granite 3.3 2B Instruct. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 2.0 GB of usable unified memory (RAM minus macOS overhead).

Can I run Granite 3.3 2B Instruct locally?

Yes — Granite 3.3 2B Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Granite 3.3 2B Instruct?

At Q4_K_M, Granite 3.3 2B Instruct can reach ~2412 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~329 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 1.99 × 0.65 = ~2613 tok/s

Estimated speed at Q4_K_M (2.0 GB)

NVIDIA B200

~2613 tok/s

NVIDIA GeForce RTX 4090

~329 tok/s

NVIDIA B300

~2613 tok/s

AMD Instinct MI350X

~2412 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Granite 3.3 2B Instruct?

At Q4_K_M, the download is about 1.52 GB. The full-precision BF16 version is 5.07 GB. The smallest option (IQ2_XXS) is 0.70 GB.

Which GPUs can run Granite 3.3 2B Instruct?

52 consumer GPUs can run Granite 3.3 2B Instruct at Q4_K_M. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT. 52 GPUs have plenty of headroom for comfortable inference.

Which devices can run Granite 3.3 2B Instruct?

163 devices with unified memory can run Granite 3.3 2B Instruct at Q4_K_M (2.0 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.