Meta · open-weights

Llama 3.2 90B

Llama 3.2 90B is an open-weights language model from Meta, released September 24, 2024. It has 90B parameters (dense). On the Epoch Capabilities Index it scores 122.8, rank 178 of 256 rated models.

Check it against your own PC →

Lab

Can you run Llama 3.2 90B locally?

The smallest common GPU it runs on is 32 GB of VRAM (with 32 GB of RAM): IQ2_XXS, fully on GPU, roughly 44 tokens per second. Estimates assume 32 GB of DDR5 RAM, a 16K context and llama.cpp-style GGUF quantization. Enter your exact hardware in the live app.

GPU memory	Verdict	Best quant	VRAM	RAM	tok/s
6 GB VRAM	Very slow	IQ2_XXS (heavy loss)	5.2 GB	22.9 GB	~1
8 GB VRAM	Very slow	IQ2_XXS (heavy loss)	7.2 GB	20.9 GB	~1
12 GB VRAM	Very slow	IQ2_XXS (heavy loss)	11.2 GB	16.9 GB	~2
16 GB VRAM	Very slow	IQ2_XXS (heavy loss)	15.2 GB	12.9 GB	~2
24 GB VRAM	Very slow	IQ2_XXS (heavy loss)	23.2 GB	4.9 GB	~5
32 GB VRAM	Runs on GPU	IQ2_XXS (heavy loss)	28.1 GB	–	~44

Strengths by category

ModelTape's 0–100 fit on Epoch AI benchmark results.

Category	Score	Based on
Math	11.6	MATH level 5, OTIS Mock AIME 2024-2025
Knowledge	29.0	GPQA diamond, MMLU

Benchmark results

Benchmark (Epoch AI)	Score
MMLU	73.7%
MATH level 5	39.4%
Balrog	27.3%
GPQA Diamond	22.3%
OTIS Mock AIME	2.5%

Links

Epoch AI benchmarks

Models with a similar score

GPT-4.1 nano · Mistral Large 2 (Jul 2024) · Llama 3.3 70B · GPT-4 Turbo (Apr 2024) · Claude 3 Opus · Gemini 1.5 Pro (May 2024) · Claude 3.5 Haiku · GPT-4 Turbo (Nov 2023)