When someone begins exploring local AI the first model family they are likely to encounter is Llama. Released by Meta and iterated upon rapidly since 2023, the Llama series has become the de facto foundation for the open-model ecosystem. The majority of fine-tunes and community tutorials and quantised builds and tool integrations are built on or optimised for Llama first. Understanding what the family offers and which variant suits which hardware and task is a practical prerequisite for anyone running inference on their own machine.
A brief history of the family
The original Llama and Llama 2 releases established the foundation but it is the Llama 3 generation that defines the current mainstream. Llama 3.1 shipped in three sizes. An 8-billion-parameter model became the standard starting point for most new local AI users. A 70-billion-parameter model aimed at users with high-end hardware. And a 405-billion-parameter model intended primarily for research and multi-GPU deployments. Each brought instruction-following and reasoning capabilities that for the first time placed a consumer-accessible model genuinely close to the quality of then-current cloud services on routine tasks.
Llama 3.2 extended the family in two directions. Smaller 1-billion and 3-billion-parameter models arrived to serve edge devices and scenarios where extremely low memory usage is the priority. More significantly, the release introduced Vision variants. The 11-billion and 90-billion multimodal models accept images as inputs alongside text, which enables tasks such as describing photographs and reading charts and analysing documents without resorting to a separate specialised model.
Llama 3.3 70Bsubsequently refined the 70-billion tier substantially. Meta's own evaluations, broadly echoed by community benchmarks, found that the 3.3 70B delivers output quality approaching that of the far larger 405B model while requiring a fraction of the resources. If you have hardware capable of running the 70B class, this release is generally considered the current sweet spot between capability and practicality.
The most recent generation, Llama 4, shifts to a mixture-of-experts (MoE) architecture. Scout carries 109 billion parameters in total, of which only 17 billion are active during any given inference step, alongside a very long context window suited to document-heavy workloads. Maverick scales the same active-parameter count to 17 billion but with 400 billion total parameters and targets higher reasoning quality for multi-GPU deployments. The MoE design means the effective compute cost during inference is much lower than the headline parameter count implies, though the full model weights must still reside somewhere accessible.
The licence
Llama models are released under the Meta Llama Community Licence, which is permissive for a wide range of personal and commercial uses. It is however a bespoke licence rather than a fully open one such as Apache 2.0. It includes restrictions on using the outputs to train competing models and imposes obligations on applications that reach significant user scales. For the overwhelming majority of local users the licence is of no practical consequence. That includes individuals running a model for personal productivity and developers building internal tools and researchers conducting experiments. It is nonetheless worth reviewing before distributing a product built upon the weights.
Why Llama dominates the ecosystem
Raw model quality is important but does not fully explain Llama's position. The deeper reason is the compounding advantage of ecosystem size. Llama attracted the largest early community, so it accumulated the greatest number of fine-tuned variants. Those are models adapted for specific tasks such as coding and creative writing and medical question answering and roleplay. Every major inference runtime including Ollama and llama.cpp and LM Studio optimised for Llama first. Quantisation builds appear sooner and in more format variants. Tutorials are written with Llama as the worked example. This accumulation means that a user choosing Llama encounters fewer rough edges than a user choosing a less-represented family even when the alternative model has comparable raw capability.
The practical consequence is that Llama is the lowest-friction entry point for local AI. Support is broad and documentation is plentiful, and problems you are likely to encounter have usually already been solved by someone in the community.
What Llama is well suited for
Llama 3.x models are general-purpose assistants. They perform creditably across a wide range of tasks without specialisation:
- General chat and writing assistance. Drafting and editing and paraphrasing and brainstorming are all comfortable territory for models of 8 billion parameters and above.
- Summarisation and document Q&A. The models handle retrieval-augmented generation (RAG) pipelines well, where relevant excerpts from a document are supplied as context and the model is asked to reason over them. This is a practical approach for querying private documents without ever sending them to a cloud service.
- A base for fine-tuning. The weights are openly available, so Llama is the most common starting point for users who wish to train a specialised model on a custom dataset.
- Vision tasks. The Llama 3.2 Vision variants (11B and 90B) add image understanding, which broadens the range of pipelines you can build locally without switching to a different family.
- Long-context workloads.Llama 4 Scout's extended context window makes it a candidate for workflows involving lengthy documents or extended conversations, provided the hardware can accommodate the model.
Hardware considerations
The 8B model is the natural starting point for anyone new to local AI. It fits within the VRAM available on a wide range of consumer graphics cards and responds at a comfortable speed on most mid-range and recent hardware. It is a capable assistant for everyday tasks and it is the model most tutorials and guides assume by default.
The 70B class including Llama 3.3 70B requires substantially more video memory. A single high-end consumer card with 24 GB of VRAM can run it with careful quantisation, often with some layers offloaded to system RAM to manage the fit. For fully in-VRAM inference a professional card or a multi-GPU setup is generally necessary. The quality difference over the 8B is meaningful on complex tasks, which is why this tier has become the target for serious local users rather than the 405B or Llama 4 Maverick, both of which demand multi-GPU or server-class deployments.
For a precise assessment of which Llama variants a specific graphics card can run and at what speed, the WillMyGPURunIt calculator provides estimates based on the actual hardware in a build. The VRAM requirements guide explains the underlying relationship between parameter count and memory, and the best GPUs for local LLMs ranks current cards by their suitability for this workload.
Llama in context
Llama is not the only compelling family available for local use. Qwen and Mistral and Gemma and DeepSeek each have their strengths, and a thorough comparison is available in the best local AI models guide. For most users Llama represents the lowest-risk starting point. The ecosystem backing it is the largest, the tooling support is the most mature, and the 8B model offers a useful capability baseline on accessible hardware. Users who outgrow the 8B and possess the hardware for it will find the 3.3 70B a substantial step up without requiring a change of toolchain or runtime.
The decision of which specific Llama variant to run ultimately comes down to VRAM. Establishing what the available graphics card can accommodate, either by consulting the guides above or entering a build into the calculator, is the most useful first step before committing to any particular model.