Llama 3.1¶

Development line: project:llama-3-1 · thread llama-3-1
Last event: 2024-07-23 · 1 dated since 2024-07-23 · Researched: 2026-09-04 · confidence: medium

What it is¶

Llama 3.1 — Meta’s static 2024 family for teams that need a downloadable, licensed alternative to a hosted text-generation API.

• 8B, 70B, and 405B pretrained and Instruct variants for text/code generation, multilingual chat, RAG, and tool-enabled systems.
• 128K context and eight supported languages.
• Text in/text out, December 2023 knowledge cutoff; 405B is a multi-GPU deployment problem.

Development line¶

• 2024-07-23 — Official Llama 3.1 access resources became available. On 2024-07-23, official Llama 3.1 resources opened paths to model downloads, Meta AI, an agentic-system repository, and a hosted chat model. This marked the line's practical availability to developers and users across those channels.

What changed¶

• 2024-07-23 — Llama 3.1 released upgraded 8B and 70B models plus 405B, with 128K context, eight supported languages, tool-use support, and companion safety/system components.
• 2024-09-25 — Llama 3.2 added 1B/3B edge text models and 11B/90B vision models; Llama 3.1 remained text-only.
• 2024-12-06 — Llama 3.3 introduced a later 70B Instruct text model with 128K context.
• 2025-04-05 — Llama 4 Scout and Maverick introduced Meta’s newer native-multimodal mixture-of-experts Llama generation.

How to use this¶

As of 2024-07-23, teams could evaluate Llama 3.1 through official download, chat, and developer channels; confirm the exact model variant and terms from official materials before deployment.

1. Confirm that text-only input/output, the 8B/70B/405B size range, and eight-language support fit the workload; choose an Instruct checkpoint for assistant chat. — https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md
2. Request gated access and accept the Llama 3.1 Community License before downloading model weights. — https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
3. Run a first local chat with Transformers 4.43+ using the official Instruct model ID and a text-generation pipeline or AutoModelForCausalLM. — https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
4. Send role-based messages through tokenizer.apply_chat_template before generation rather than assembling Llama control tokens yourself. — https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
5. Serve the checkpoint with vLLM and call /v1/chat/completions for an OpenAI-compatible local endpoint. — https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
6. Plan a node with 8×80GB H100 GPUs for 405B FP8; use multi-node inference for the unquantized 405B checkpoint. — https://github.com/meta-llama/llama-cookbook/blob/main/getting-started/inference/local_inference/README.md

Best practices¶

• Use an Instruct model for assistant dialogue; keep the pretrained base model for adaptation or other text generation. — https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md
• Preserve the official chat template with apply_chat_template, including the generation prompt. — https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
• Follow the Llama 3.1 Community License as a release requirement: distributed materials and services have attribution obligations, and products above 700 million monthly active users require a separate Meta license. — https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct-FP8
• Safety-test and tune for the application, then deploy safeguards around the model and external tools; do not deploy the model alone. — https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md
• Do not budget 405B from its parameter label alone: separate FP8 single-node serving from unquantized multi-node serving. — https://github.com/meta-llama/llama-cookbook/blob/main/getting-started/inference/local_inference/README.md

Superseded by this¶

• 2024-09-25 — Choosing Llama 3.1 for on-device text or image understanding is obsolete; Llama 3.2 introduced the 1B/3B edge-text and 11B/90B vision paths.
• 2024-12-06 — Comparing only Llama 3.1 70B or 405B for a new Meta 70B text service is obsolete; Llama 3.3 70B Instruct is the later 128K option.
• 2025-04-05 — Calling Llama 3.1 Meta’s newest Llama generation is obsolete; Llama 4 added native multimodality and mixture-of-experts models.

Still unknown¶

• The source and X links yielded no retrievable text in this pass, so we use only their dates and URLs rather than unverified claims.
• The historical llama-agentic-system link redirects to ogx-ai/llama-stack-apps; that page lacks dated Meta provenance, so we do not treat it as current official Meta guidance.
• We found a same-day Chinese Hugging Face translation, but no separate primary Meta announcement; we make no China-specific availability or compatibility claims.
• We have no workload, hardware budget, target region, or license context, so we do not pick a specific Llama 3.1 size or later replacement.

Sources¶

Agent brief¶

• Subject: project:llama-3-1, thread llama-3-1, 1 dated events 2024-07-23 → 2024-07-23.
• Practical note: As of 2024-07-23, practitioners could evaluate Llama 3.1 through official download, chat, and developer-resource channels; confirm the exact model variant and applicable terms from the official materials before deployment.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.