Llama-3.3-70B-Instruct

Meta’s Llama 3.3 70B Instruct is a 70-billion-parameter, pretrained and instruction-tuned generative language model designed for multilingual dialogue. It supports English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai, making it well suited to assistants and text-generation applications that need one model across several languages. Its instruction tuning is intended to produce responsive, assistant-style behavior for general conversational workflows.

The model is positioned for stronger reasoning, mathematics, and instruction following, while its reported evaluation profile includes graduate-level scientific reasoning, instruction-following, and agentic terminal tasks. It also supports function calling and structured JSON output, adding practical utility for tool-connected systems and applications that need predictable response formatting. Its broad language coverage and combination of open-weight availability, dialogue optimization, and tool-oriented features make it a versatile choice for multilingual text assistants, though actual deployment performance will depend on the chosen serving environment.

watsonx.aimeta-llama/llama-3-3-70b-instructllama

Open provider page Provider docs

Quick Info

Provider: watsonx.ai
Model key: meta-llama/llama-3-3-70b-instruct
Release date: Dec 6, 2024
Last updated: Dec 6, 2024
Knowledge cutoff: 2023-12
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.7526
Output token cost: $0.7526

Limits

Output tokens: 4,096 tokens
Context window: 131,072 tokens

Transparent token rates

Compare Llama-3.3-70B-Instruct pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Llama-3.3-70B-Instruct

Kilo Gateway

Coverage

A D-Central model profile documents Llama 3.3 as a single 70B dense transformer released by Meta in December 2024, positioning it as a post-training refresh of the Llama 3.1 70B backbone rather than a from-scratch retrain. The page credits Meta as developer, lists the model under the Llama family with the Llama 3.3 Com Technical details on the profile show the architecture is unchanged from 3.1 70B — same 80 layers, 8192 hidden dim, SwiGLU activations, RoPE positional encoding, and Grouped-Query Attention — with the gains attributed to an improved instruction-tuning pipeline: better RLHF datasets, a stronger reward model, and an adde