Llama model series
Meta's Llama line from Llama 1 to Llama 4: release months, sizes, context length, licenses, and which generation to use today for which job.
What Llama is
Llama is Meta's open-weight language model line. It is the most widely fine-tuned base in open models, and many derivatives on Hugging Face trace back to one of the generations below. Every Llama release uses a custom Meta community license, not Apache 2.0 or MIT, so read the terms for your generation before shipping a product. Each generation has its own hub listing the models we track for it.
Generations in order
Llama 1 (February 2023). Meta announced LLaMA on February 24, 2023 in four sizes: 7B, 13B, 33B and 65B. Access was limited to researchers under a noncommercial license. The weights most people use today are community re-uploads such as huggyllama/llama-7b, which appeared in April 2023.
Llama 2 (July 2023). Sizes 7B, 13B and 70B, with a 4,096 token context, trained on 2.0 trillion tokens. The Llama 2 Community License permits commercial use, which is why this generation spread so far.
Llama 3 (April 2024). Released April 18, 2024. The 8B Instruct card lists an 8,192 token context, grouped-query attention, and training on more than 15 trillion tokens. A 70B model sits beside it.
Llama 3.1 (July 2024). Released July 23, 2024 in 8B, 70B and 405B sizes, with the context raised to 128,000 tokens and support for eight languages including German, French, Hindi and Thai.
Llama 3.2 (September 2024). Released September 25, 2024. The small text models are 1B (1.23B parameters) and 3B (3.21B parameters), with 128k context in the full-precision versions and 8k in the quantized ones.
Llama 3.3 (December 2024). One model: Llama 3.3 70B Instruct, released December 6, 2024, with a 128,000 token context and the same eight languages.
Llama 4 (April 2025). The switch to mixture-of-experts and native multimodality (text and image input, early fusion). Both released April 5, 2025:
- Llama 4 Scout: 17B active, 109B total, 16 experts, 10 million token context, trained on about 40 trillion tokens.
- Llama 4 Maverick: 17B active, 400B total, 128 experts, 1 million token context, trained on about 22 trillion tokens.
The Llama 4 license requires a separate agreement for products with more than 700 million monthly active users.
Which generation to use today
Meta's Hugging Face organization still lists Llama 4 and Llama 3.3 as its current text models, so we found no newer Meta generation to point you to.
- Long context or multimodal input: Llama 4 Scout has the longest context in the line (10 million tokens) and takes images. All 109B weights must fit in memory even though 17B are active.
- One 70B-class text model: Llama 3.3 70B Instruct, dense, 128,000 token context.
- Small models: Llama 3.2 1B and 3B, or Llama 3.1 8B, which also has a 128,000 token context.
- Fine-tuning from a familiar base: Llama 3.1 and 3 have the largest ecosystem of derivatives, but their licenses still apply to what you build.
- Avoid for new work: Llama 1 (research-only terms) and Llama 2 (4,096 token context) unless you need to reproduce older results.
Reading the licenses
Every generation carries a different Meta license, named after the version: the Llama 2 Community License, the Meta Llama 3 Community License, the Llama 3.1, 3.2 and 3.3 Community Licenses, and the Llama 4 Community License Agreement. The gated repositories ask for contact information before access. These are custom licenses, so check the text for the exact generation before redistributing or building a derivative.
Side lines
Code Llama is the code-specialized branch. Llama Guard holds the safety classifiers: Llama Guard 4 is a 12B model published in April 2025, and the Meta organization also lists Prompt Guard models at 86M and 22M parameters.
Llama models can run on Aquanode GPUs.
Llama generations
Every Llama generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Llama 1 | 1 | 6.7B | 15.1 GB | V100 | $0.110/hr |
| Llama 2 | 8 | 6.7B to 13.0B | 15.1 GB | V100 | $0.110/hr |
| Llama 3 | 19 | 8.0B to 70.6B | 17.9 GB | RTX A5000 | $0.176/hr |
| Llama 3.1 | 40 | 950M to 405.9B | 17.9 GB | RTX A5000 | $0.176/hr |
| Llama 3.2 | 15 | 506M to 88.6B | 2.8 GB | RTX 3070 | $0.088/hr |
| Llama 3.3 | 4 | 49.9B to 70.6B | 158 GB | AMD MI300X | $2.63/hr |
| Llama 4 | 2 | 108.6B to 401.6B | 243 GB | RTX A6000 × 6 | $2.18/hr |
Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.
Other Llama lines
Side lines from Meta: specialised or companion models outside the main Llama generations.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Code Llama | 3 | 6.7B to 33.7B | 15.1 GB | RTX 5060 Ti | $0.145/hr |
| Llama Guard | 4 | 1.2B to 12.0B | 3.3 GB | RTX 3070 | $0.088/hr |
Run and fine-tune Llama
Per-generation guides: engine support, chat template and context flags, then fine-tune memory for LoRA and QLoRA.
- Llama 4: how to run, fine-tune
- Llama 3.3: how to run, fine-tune
- Llama 3.2: how to run, fine-tune
- Llama 3.1: how to run, fine-tune
- Llama 3: how to run, fine-tune
- Llama 2: how to run, fine-tune
Llama compared with other models
Side-by-side VRAM, context length and license for models of a similar size.
- Llama-3.2-1B-Instruct vs Qwen3-0.6B
- Llama-3.2-3B-Instruct vs Qwen2.5-3B-Instruct
- Llama-3.1-8B-Instruct vs Qwen3-8B
- Llama-3.3-70B-Instruct vs Qwen-72B
- Meta-Llama-3-8B-Instruct vs Qwen3-8B
- Llama-3.1-70B-Instruct vs Qwen-72B
- Meta-Llama-3-70B vs Qwen-72B
- DeepSeek-R1 vs Llama-3.1-405B
- Llama-3.2-1B-Instruct vs Qwen2.5-1.5B-Instruct
- Llama-3.2-3B-Instruct vs Qwen3-4B
- Llama-3.1-8B-Instruct vs Qwen2.5-7B-Instruct
- Llama-3.3-70B-Instruct vs Qwen3-Coder-Next
- Meta-Llama-3-8B-Instruct vs Qwen2.5-7B-Instruct
- Llama-3.3-70B-Instruct vs Qwen2.5-72B-Instruct
- Llama-3.3-70B-Instruct vs Phi-3.5-MoE-instruct
- Llama-3.3-70B-Instruct vs Qwen2.5-72B
- DeepSeek-V3.2 vs Llama-3.1-405B
- Llama-3.2-1B-Instruct vs Qwen2.5-0.5B-Instruct
- Llama-3.2-3B-Instruct vs Qwen3-4B-Instruct-2507
- Llama-3.1-8B-Instruct vs Qwen2.5-Coder-7B-Instruct
- Llama-3.1-70B-Instruct vs Qwen3-Coder-Next
- Llama-3.2-1B-Instruct vs Qwen3-1.7B
- Llama-3.1-8B-Instruct vs Mistral-7B-Instruct-v0.2
- Llama-3.1-70B-Instruct vs Qwen2.5-72B-Instruct
- gemma-3-1b-it vs Llama-3.2-1B-Instruct
- granite-4.1-8b vs Llama-3.1-8B-Instruct
- Llama-3.1-70B-Instruct vs Phi-3.5-MoE-instruct
- Llama-3.2-1B-Instruct vs SmolLM2-135M
- DeepSeek-R1-0528-Qwen3-8B vs Llama-3.1-8B-Instruct
- Llama-3.1-70B-Instruct vs Qwen2.5-72B
- Llama-3.2-3B-Instruct vs Qwen3-4B-Base
- Meta-Llama-3-8B-Instruct vs Qwen2.5-Coder-7B-Instruct
- Meta-Llama-3-70B vs Qwen3-Coder-Next
- Meta-Llama-3-70B-Instruct vs Qwen3-Coder-Next
- DeepSeek-V3-0324 vs Llama-3.1-405B
- Llama-3.1-70B vs Qwen3-Coder-Next
- Llama-3.2-3B-Instruct vs SmolLM3-3B-Base
- Meta-Llama-3-8B-Instruct vs Mistral-7B-Instruct-v0.2
- Llama-3.2-3B-Instruct vs Phi-3-mini-4k-instruct
- granite-4.1-8b vs Meta-Llama-3-8B-Instruct
Best models by task
Llama models appear on these ranked task pages.
Sources
- https://ai.meta.com/blog/large-language-model-llama-meta-ai/
- https://huggingface.co/meta-llama
- https://huggingface.co/api/models/huggyllama/llama-7b
- https://huggingface.co/meta-llama/Llama-2-7b-hf
- https://huggingface.co/api/models/meta-llama/Llama-2-7b-hf
- https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct
- https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
- https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
- https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct
- https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct
- https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct
Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.
Related on Aquanode: GPU requirements for CodeLlama-34b-Instruct-hf, GPU requirements for Llama-2-13b-hf, GPU requirements for Meta-Llama-3-70B, GPU requirements for Llama3-ChatQA-1.5-8B, and GPU requirements for Llama-3.1-70B.