Llama 3.1 8B Instruct
The 8B instruction-tuned model of Meta's Llama 3.1 collection, for chat and code in eight languages. It was pretrained on roughly 15 trillion tokens, and its knowledge stops at December 2023.
Overview
Llama 3.1 8B Instruct is the smallest instruction-tuned model in Meta's Llama 3.1 collection, released on July 23, 2024 alongside 70B and 405B versions. Meta pretrained it on roughly 15 trillion tokens of publicly available data, and post-training combined supervised fine-tuning with RLHF, using public instruction sets plus over 25 million synthetic examples. Its knowledge stops at December 2023. It supports English, German, French, Italian, Portuguese, Hindi, Spanish and Thai, and it also writes code.
Meta reports 69.4 on MMLU (5-shot), 80.4 on IFEval, 72.6 on HumanEval and 84.5 on GSM-8K with chain of thought.
The checkpoint has a 131,072-token (128K) window. The input Wiro accepts per request is smaller and has not been measured, so plan for well under 128K, counting system message, stored history and prompt together. Over-long requests are shortened, the oldest history first and then the end of your newest message. If the system message leaves no room for your message, the run fails with a context-length error. With the token limit at 0, a reply stops at 2,048 tokens for any prompt up to that length.
Choosing a variant: of the four, this is the largest instruction-tuned model and the general pick for multilingual chat. Llama 3.2 3B Instruct is smaller and adds function tools and JSON output. Llama 2 7B Chat and Llama 2 13B belong to the 2023 generation: English only, 4k context.
What you can build
- Support chat that answers in any of the eight supported languages
- Code explanation and small helper functions
- Synthetic training data for smaller models, a use the Llama 3.1 license allows
Inputs
A prompt is required. The system prompt field comes prefilled with a paragraph of safety instructions; replace it with your own role and reply language. Add a user ID, and a session ID per thread, for Wiro to store and reuse conversation history.
Temperature defaults to 0.7. Other settings: top-p, top-k, penalties, token limits, stop sequences, a seed, a sampling toggle and quantization (on by default). There are no tool or response-format inputs.
Outputs
Text, with code when asked. The sample lists eight numbered events, from the 1215 siege to the Great Fire of 1666, each paired with a story angle.
Recommended settings
- Put the instruction at the start of a long prompt, so trimming removes background text and leaves the task intact.
- Switch sampling off for code and data extraction.
Limitations
- Historical details need checking: Thomas More was beheaded on Tower Hill, outside the walls, yet the sample puts his execution on the Tower grounds.
- Meta warns it can still give inaccurate or objectionable answers and recommends safeguards such as Llama Guard 3 and Prompt Guard.
Safety & compliance
It ships under the Llama 3.1 Community License. Redistributors must include the license and prominently display "Built with Llama", and a distributed model trained on its outputs must start its name with "Llama". Meta's Acceptable Use Policy prohibits illegal activity, harassment, malware and disinformation, among others. Languages outside the eight listed are out of scope unless you fine-tune for them under the license terms.
API quick start
Run Llama-3.1-8B-Instruct with a single API call.
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/meta-llama/llama-3-1-8b-instruct" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.



