# Llama 3.2 3B Instruct ![Organization](https://img.shields.io/badge/Org-Meta-0064e0) ![Parameters](https://img.shields.io/badge/Params-3B-1f6feb) ![Context](https://img.shields.io/badge/Context-128K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Llama%203.2-d29922) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Sep%202024-555) > **Quick answer:** Llama 3.2 3B Instruct is Meta's September 2024 small open-weight model, scoring 77.7% on GSM8K, 69.8% on HellaSwag, and 67.0% on BFCL v2. With a 128K context window and Llama 3.2 License, it is designed for on-device deployment and edge inference. ## At a Glance **Where Llama 3.2 3B leads** - 77.7% GSM8K — strong math reasoning at 3B scale - 128K context window — unusually long for a 3B model - On-device capable — fits mobile and edge hardware - 67.0% BFCL v2 — solid tool use at 3B scale - Llama 3.2 License permits on-device distribution - Strong foundation for 3B-scale fine-tuning **Where it lags** - 69.8% HellaSwag — common-sense reasoning below larger models - Text-only in the 3B variant (Llama 3.2 11B/90B add vision) - Significantly below 8B+ models on all benchmarks - December 2023 knowledge cutoff (older than 70B variant's March 2024) **Best for:** On-device inference; mobile applications; edge deployments; rapid response applications requiring low latency at 3B scale. ## What Llama 3.2 3B Instruct Is Llama 3.2 3B Instruct (released September 25, 2024) is Meta's smallest instruction-tuned model in the Llama 3.2 family. Its primary design goal is on-device and edge deployment — running on mobile phones, laptops, and embedded hardware without cloud inference. Despite its small size, Llama 3.2 3B achieves 77.7% GSM8K — a score that would have been competitive with state-of-the-art models just two years earlier. Its 128K context window is exceptionally long for a 3B model, enabling long-document tasks at the edge. The Llama 3.2 3B is part of Meta's broader push for on-device AI with the Llama 3.2 family. The 1B variant goes even smaller for more constrained hardware; the 11B and 90B variants in the same family add multimodal vision capability. ## Specifications | Field | Value | |---|---| | Organization | Meta | | Parameters | 3B (dense) | | Context window | 128,000 tokens | | License | Llama 3.2 License | | HuggingFace | [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) | | Release date | September 25, 2024 | | Knowledge cutoff | December 2023 | | Modality | Text only | ## Pricing Llama 3.2 3B Instruct is available as open weights on Hugging Face under the Llama 3.2 License. Available via numerous hosted providers at extremely low cost due to small model size. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [GSM8K](/benchmarks/gsm8k) | 77.7% | Benchgen evaluation | 2025-07 | | [HellaSwag](/benchmarks/hellaswag) | 69.8% | Benchgen evaluation | 2025-07 | | [BFCL v2](/benchmarks/bfcl-v3) | 67.0% | Benchgen evaluation | 2025-07 | ## Llama 3.2 3B vs Alternatives | Model | GSM8K | HellaSwag | BFCL v2 | Params | |---|---|---|---|---| | **Llama 3.2 3B Instruct** | 77.7% | 69.8% | 67.0% | 3B | | [Llama 3.1 8B Instruct](/models/meta/llama-3-1-8b-instruct) | 84.5% | — | — | 8B | | [Phi-4](/models/microsoft/phi-4) | — | — | — | 14B | Llama 3.2 3B is designed for on-device use where the 8B model won't fit. For server-side deployments, Llama 3.1 8B Instruct provides meaningfully better scores (84.5% vs 77.7% GSM8K) at modest cost increase. ## Related Models & Benchmarks - Compare: [Llama 3.1 8B Instruct](/models/meta/llama-3-1-8b-instruct) · [Llama 3.1 70B Instruct](/models/meta/llama-3-1-70b-instruct) - Benchmarks: [GSM8K](/benchmarks/gsm8k) · [HellaSwag](/benchmarks/hellaswag) · [BFCL v3](/benchmarks/bfcl-v3) ## Frequently Asked Questions

What is Llama 3.2 3B Instruct?

Llama 3.2 3B Instruct is Meta's September 2024 small on-device model scoring 77.7% GSM8K, 69.8% HellaSwag, and 67.0% BFCL v2 with a 128K context window.

Can Llama 3.2 3B run on-device?

Yes — that is its primary design goal. At 3B parameters (quantised to 4-bit, ~1.8GB), it fits on most modern smartphones and edge hardware.

Should I use Llama 3.2 3B or Llama 3.1 8B?

Use 3B for on-device or edge deployment where 8B won't fit. For server-side inference, Llama 3.1 8B Instruct provides significantly better benchmark performance (84.5% vs 77.7% GSM8K) at low cost.

--- *Specs from Meta's official Llama 3.2 release (September 2024) and Benchgen evaluations. Last updated 2026-07-24.*

Llama 3.2 3B Instruct — Files