Llama 3.2

Claims and evidence

• Released 25 September 2024 ● Primary source[1]page date; Takeaways [3]Model Release Date [5]README, Llama Models table

• Status AvailableMeta publishes no deprecation page for Llama models; the weights of all four sizes and the quantized 1B and 3B builds are still listed in Meta's Hugging Face organisation (gated, license acceptance required) as seen on 2026-10-01. ● Primary source[4]

• Derived from (distillation) Llama 3.1The 1B and 3B models were pruned from Llama 3.1 8B and trained with logits from Llama 3.1 8B and 70B as token-level targets. The 11B and 90B Vision models start from pretrained Llama 3.1 text models and add a separately trained image adapter and encoder, leaving the language model weights unchanged. ● Primary source[1]Lightweight models; Vision models [2]Training Data [3]Model Architecture
• Successor of Llama 3.1 EditorialEditorial link along the Llama 3 version line: the announcement presents Llama 3.2 as the release after the Llama 3.1 herd, without calling it a successor. ● Primary source[1]

• Change · Modality Adds image input in the 11B and 90B Vision models; for image plus text tasks English is the only supported language.Compared with Llama 3.1 ● Primary source[1]Vision models [3]Model Information
• Change · Size Adds 1B and 3B text models for mobile and edge devices, enabled for Qualcomm and MediaTek hardware and optimised for Arm processors.Compared with Llama 3.1 ● Primary source[1]Takeaways; Lightweight models
• Change · Efficiency Quantized builds of the 1B and 3B Instruct models followed on 24 October 2024, limited to 8K context for mobile devices.Compared with Llama 3.1 ● Primary source[6]
• Change · Licensing The acceptable use policy does not grant the license rights for the multimodal models to individuals or companies based in the European Union.Compared with Llama 3.1 ● Primary source[7]paragraph on multimodal models after the list of prohibited uses

• Input textApplies to the 1B and 3B text models; the two Vision variants also accept images. ● Primary source[2]Model Information table, Input modalities
• Output text, code ● Primary source[2]Model Information table, Output modalities
• Open weights YesWeights under the custom Llama 3.2 Community License, after accepting the license. The acceptable use policy does not grant the license rights for the multimodal models to individuals or companies based in the European Union; end users of products that include them are exempt. ● Primary source[1]Takeaways [4] [7]paragraph on multimodal models after the list of prohibited uses
• Context window 128K tokensThe quantized 1B and 3B builds of October 2024 are limited to 8k. ● Primary source[1]Takeaways (1B and 3B) [2]Model Information table, Context Length (128k) [3]Model Information table, Context length (128k)
• Access Open-weights download, cloud partner, On device, consumer appOn-device: the 1B and 3B models (and their quantized builds) target mobile and edge hardware. Consumer access: Meta says the Vision models can be tried in Meta AI. ● Primary source[1]Takeaways [6]Takeaways

Lineage

Variants

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Meta on AI Radar →