Answer · confidence: partial
When could people first run AI at home, and what does it need?
Research question: When did AI models first become available for people to download and run on their own computers, and what hardware do they need?
The records show that Meta's first Llama model had its inference code publicly released under the GPLv3 license on February 24, 2023 2, but its weights were gated behind an application process 3, and unauthorized copies later spread via BitTorrent 4. The facts do not establish that Llama was the first AI model people could download and run, so the question of the very first release is only partly answered 23. For hardware, a 4-bit quantized Llama 3.x 8B model can run on consumer hardware such as an RTX 4060 Ti 16GB or a 16 GB Apple Silicon Mac 7, and a rule of thumb puts a 7B 4-bit model at roughly 4-5 GB of VRAM 14. NVIDIA GPUs are strongly recommended for Llama 3 because the primary tooling targets CUDA 10, while the much larger DeepSeek V3 and R1 models have 671 billion total parameters 11. The newest fact is dated October 17, 2026 14, so it is current as of today.
Checked 2026-10-20. Only documented and supported facts are used.
Dig deeper on this questionSources
[1] supported
Open weights can support local deployment, customization, and independent research without requiring access through the original developer's API.
"Open weights can support local deployment, customization, independent research and competition without requiring access through the original developer's API ."Open weights
[2] supported
The first Llama model's inference code, which is the code needed to run the model, was publicly released under the open-source GPLv3 license.
"The inference code used to run the model was publicly released under the open-source GPLv3 license."Llama (language model)
[3] supported
Access to the first Llama model's weights, the trained parameters needed to run it, was controlled through an application process rather than being openly downloadable.
"Access to the model's weights was managed by an application process"Llama (language model)
[4] supported
Unauthorized copies of the first Llama model were shared via BitTorrent, letting people get it outside the approved access process.
"Unauthorized copies of the first model were shared via BitTorrent"Llama (language model)
[5] supported
Llama was trained at several model sizes with the stated intention of making it more accessible to different hardware.
"Llama was trained at various model sizes, with the intention to make it more accessible to different hardware."Llama (language model)
[6] supported
Llama 2 was released in three sizes (7, 13, and 70 billion parameters) with model weights made available, and was usable for many commercial use cases.
"Meta trained and released Llama 2 in three model sizes: 7, 13, and 70 billion parameters."Llama (language model)
[7] supported
Llama 3.x 8B models at 4-bit quantization can run on consumer hardware such as an RTX 4060 Ti 16GB or an Apple Silicon Mac with 16 GB of unified memory.
"For Llama 3.x 8B at 4-bit quantisation, 8 GB of VRAM is enough. A consumer card like the RTX 4060 Ti 16GB or an Apple Silicon Mac with 16 GB unified memory works."Run Llama 3 Locally: Hardware, Tooling, and Setup
[8] supported
Meta's Llama 3 family of open-weight large language models was first released in April 2024.
"Llama 3 is Meta's family of open-weight large language models, first released in April 2024 and developed by Meta AI."Llama 3 Hardware Requirements & GPU Guide
[9] supported
Because Llama 3 is released as open weights, it can be downloaded and run on owned hardware for private inference without API fees.
"Because Llama 3 is released as open weights under a community license, the models can be downloaded and run on owned hardware for fully private inference, fine-tuning, and retrieval-augmented generation, with no API fees and no data leaving the machine."Llama 3 Hardware Requirements & GPU Guide
[10] supported
NVIDIA GPUs are strongly recommended for running Llama 3 locally because the primary tooling targets CUDA.
"NVIDIA GPUs are strongly recommended because the local LLM tooling targets CUDA first."Llama 3 Hardware Requirements & GPU Guide
[11] supported
The DeepSeek V3 and R1 models have 671 billion total parameters with 37 billion active per token.
"DeepSeek V3 and R1 represent the frontier of Mixture-of-Experts language models at 671 billion total parameters (37B active per token)."Run inference (2026)
[12] supported
GPT-OSS-20B is an open-weight model that OpenAI designed to run on consumer hardware with at least 16GB of memory.
"for reasoning on consumer hardware with ≥16GB"Private LLM Inference on Consumer Blackwell GPUs: A Practical Guide for Cost-Effective Local Deployment in SMEs (2026)
[13] supported
An NVIDIA RTX 4090 consumer GPU, priced at about $2,000, is only 18% slower than a top-tier A100 GPU costing around $20,000 on the model tested.
"at approximately $2,000) is only 18% slower compared to"PowerInfer: Fast Large Language Model Serving with a Consumer-grade GPU (2024)
[14] supported
A common rule of thumb is that 4-bit quantized models need about 0.6 GB of VRAM per billion parameters, so a 7B model needs roughly 4-5 GB.
"Rule of thumb for 4-bit quantization: VRAM in GB ≈ parameters in billions × 0.6 — a 7B model needs roughly 4-5 GB."Local LLM Hardware Requirements: 8GB to 70B by VRAM