llama nsfw on Anyken: compare the model examples with your intended input and output. Review the capabilities documented below, then use the controls actually offered in the workspace. Available settings and credits are shown before submission.
Read source
This section provides an overview of the Llama model series as documented in the official repository, covering model versions, key features, and the open-access philosophy.
The Llama model series, developed by Meta, includes multiple versions released over time. According to the official GitHub repository, Llama 2 launched with 7B, 13B, and 70B parameter sizes and a 4K context length. Llama 3 followed with 8B and 70B variants and an 8K context length. Llama 3.1 expanded to include a 405B parameter model with a 128K context length. Llama 3.2 introduced 1B and 3B models, also with 128K context, alongside vision-capable variants at 11B and 90B. Llama 3.3 offered a 70B model with 128K context. Llama 4, released in April 2025, includes Scout-17B-16E and Maverick-17B-128E configurations with context lengths of 10M and 1M tokens respectively. All models use tokenizers appropriate to their generation, with earlier versions using SentencePiece and later versions adopting a TikToken-based tokenizer. The repository provides model cards, acceptable use policies, and licenses for each version.
The Llama model series is described as an accessible, open large language model designed for developers, researchers, and businesses. Key aspects highlighted in the documentation include open access to cutting-edge models, a broad ecosystem with hundreds of millions of downloads and thousands of community projects, and a comprehensive approach to trust and safety. Llama nsfw weights are licensed for both researchers and commercial entities, with the stated mission to empower individuals and industry while fostering ethical AI advancements. The repository notes that platform support is broad, ranging from cloud providers to startups. The documentation emphasizes that Llama models serve as a foundational system for innovation in the global community. The acceptable use policy and license for each model version are provided alongside the model card, which details intended uses and limitations. The repository also provides a Responsible Use Guide to help developers address potential risks associated with the technology.
This section details the technical capabilities of Llama models and the documented methods for accessing them, including download procedures and inference options.
Accessing Llama models requires following a documented process. Users must visit the Meta Llama website, read and accept the license, and upon approval receive a signed URL via email. The llama-models CLI tool facilitates downloading: after installing with 'pip install llama-models', users run 'llama-model list' to see available models and determine llama nsfw ID. The command 'llama-model download --source meta --model-id CHOSEN_MODEL_ID' initiates the download, with the user prompted to pass the URL received. The repository notes that download links expire after 24 hours and a certain number of downloads, and users can re-request a link if errors such as 403: Forbidden occur. Additional CLI commands include listing all models including older versions, showing detailed model information, verifying download integrity, removing downloaded models, and displaying prompt formats. Downloads are also available via Hugging Face, where users visit a model repository, accept the license, and gain access to all Llama 3.1 models and previous versions.
Running Llama models requires installing dependencies within a suitable Python environment using 'pip install .[torch]'. Example scripts are available in the models/llama3 and models/llama4 sub-directories. The Llama 4 series requires at least 4 GPUs for full bf16 precision inference. The repository provides a torchrun command specifying the number of GPUs, checkpoint directory, and world size. For reduced memory footprint, the documentation describes two quantization modes: fp8_mixed, which uses FP8 for some weights and bfloat16 for activations, and int4_mixed, which uses Int4 for some weights and bfloat16 for activations. Using FP8 quantization, Llama-4-Scout-17B-16E-Instruct requires 2 GPUs with 80GB of memory each; using Int4 quantization, a single GPU with 80GB of memory suffices. The repository also provides instructions for using Hugging Face transformers, including a code snippet for loading llama nsfw and tokenizer, applying chat templates, and generating responses with torchrun across multiple GPUs.
This section covers the documented availability of Llama models, including download sources, licensing, and platform support.
Llama model weights are available through two primary channels. The official Meta Llama website requires users to read and accept the license, after which a signed URL is provided via email for downloading. The llama-models CLI tool facilitates this process. Alternatively, downloads are available on Hugging Face in both transformers and native llama4 formats. Users visit a model repository such as meta-llama/Llama-4-Scout-17B-16E, accept the license, and upon approval gain access to all Llama 3.1 models and previous versions. The repository notes that requests used to take up to one hour to process. For native weights compatible with the repository, users can download the original folder contents from the 'Files and versions' tab or use the huggingface-cli tool. The documentation also provides a transformers-compatible snippet that downloads and caches weights automatically when running inference.
Each Llama model version is accompanied by a specific license and acceptable use policy, as documented in the repository. Llama nsfw weights are licensed for both researchers and commercial entities, upholding principles of openness. The repository provides links to the Acceptable Use Policy and License for each model version, from Llama 2 through Llama 4. The documentation emphasizes that Llama models are part of a comprehensive approach to trust and safety, releasing models and tools designed to enable community collaboration and encourage standardization of trust and safety tools for generative AI. The Responsible Use Guide is provided to help developers address potential risks. The repository also includes instructions for reporting issues with the model, reporting risky content generated by the model, and reporting bugs and security concerns through designated channels. A FAQ is available for common questions and is updated over time.
This section describes the documented configurations available for Llama models, including parameter sizes, context lengths, and quantization options.
The Llama model series offers a range of parameter sizes and context lengths across versions. Llama 2 provides 7B, 13B, and 70B parameters with 4K context. Llama 3 offers 8B and 70B with 8K context. Llama 3.1 expands to 8B, 70B, and 405B with 128K context. Llama 3.2 includes 1B and 3B text models and 11B and 90B vision models, all with 128K context. Llama 3.3 provides a 70B model with 128K context. Llama 4 introduces Scout-17B-16E and Maverick-17B-128E configurations with context lengths of 10M and 1M tokens respectively. The tokenizer types also vary: Llama 2 uses SentencePiece, while Llama 3 and later versions use a TikToken-based tokenizer. Each model version has a specific launch date documented in the repository, along with links to its model card, acceptable use policy, and license.
The repository documents specific quantization modes to reduce memory footprint while maintaining accuracy. Two modes are available: fp8_mixed, which uses FP8 for some weights and bfloat16 for activations, and int4_mixed, which uses Int4 for some weights and bfloat16 for activations. For Llama-4-Scout-17B-16E-Instruct, using FP8 quantization requires 2 GPUs with 80GB of memory each, while Int4 quantization requires a single GPU with 80GB of memory. Without quantization, the Llama 4 series requires at least 4 GPUs for full bf16 precision inference. The repository provides example torchrun commands specifying the number of GPUs, checkpoint directory, and quantization mode. These configurations allow users to balance computational resources against model performance. The documentation also notes that for more flexibility in running inference, including using other providers, users should refer to the Llama Stack toolset.
This section provides practical guidance for developers working with Llama models, based on documented procedures and tools.
The llama-models CLI provides several commands for managing models. The 'llama-model list' command shows available models, with '--show-all' flag to include older versions. 'llama-model describe -m MODEL_ID' displays detailed information about a specific model. 'llama-model download' initiates downloads from Meta or Hugging Face. 'llama-model verify-download' checks the integrity of downloaded models. 'llama-model remove -m MODEL_ID' removes a downloaded model. 'llama-model prompt-format -m MODEL_ID' shows the prompt format for a model. For detailed help on any command, users can run 'llama-model COMMAND --help'. The repository also provides example scripts in the models/llama3 and models/llama4 sub-directories for running chat completion and base model completion. These scripts use torchrun for distributed inference across multiple GPUs.
The repository emphasizes that Llama models are a new technology carrying potential risks with use, and that testing conducted to date has not covered all scenarios. To help developers address these risks, the Responsible Use Guide is provided. The documentation includes specific channels for reporting issues: software bugs or other problems with the models can be reported through the GitHub issues page; risky content generated by llama nsfw can be reported at developers.facebook.com/llama_output_feedback; and bugs and security concerns can be reported at facebook.com/whitehat/info. A FAQ is available for common questions and is updated over time. The repository also provides a comprehensive approach to trust and safety, releasing models and tools designed to enable community collaboration and encourage standardization of trust and safety tools for generative AI.
Common questions about Llama models based on documented information.
The cited technical documentation for Meta's Llama model series describes model capabilities, download procedures, and responsible use guidance, but does not establish an adult-content policy or built-in NSFW filtering. The repository provides model cards, acceptable use policies, and licenses for each model version, but these documents are not included in the cited source. Readers seeking information about content moderation or NSFW filtering should consult the current acceptable use policy and service terms directly. The documentation does provide a Responsible Use Guide and channels for reporting risky content generated by llama nsfw.
Llama models can be downloaded through two official channels. The Meta Llama website requires users to read and accept the license, after which a signed URL is provided via email. The llama-models CLI tool facilitates downloading with commands like 'llama-model download --source meta --model-id CHOSEN_MODEL_ID'. Alternatively, downloads are available on Hugging Face, where users visit a model repository, accept the license, and gain access. The repository notes that download links expire after 24 hours and a certain number of downloads, and users can re-request a link if errors occur.
Running Llama 4 models at full bf16 precision requires at least 4 GPUs. With FP8 quantization, Llama-4-Scout-17B-16E-Instruct requires 2 GPUs with 80GB of memory each. With Int4 quantization, a single GPU with 80GB of memory is sufficient. The repository provides example torchrun commands for each configuration, specifying the number of GPUs, checkpoint directory, and quantization mode. For more flexibility in running inference, users can refer to the Llama Stack toolset.
