Llama 3.3 70B Instruct

Meta released Llama 3.3 70B Instruct on December 6, 2024 under the Llama 3.3 Community License, which permits commercial use. The model is a 70B auto-regressive transformer with Grouped-Query Attention, supports a 128K-token context window, uses a BPE tokenizer from the Llama 3 family, and was pretrained on 15T+ tokens with a December 2023 knowledge cutoff, as documented on the supplied model profile page. According to Meta-attributed figures cited on the page, the 70B model matches or exceeds the larger 405B Llama 3.1 on several benchmarks, including IFEval (92.1 vs 88.6), MATH CoT 0-shot (77.0 vs 73.8), and GPQA Diamond (50.5 vs 49.0), while remaining single-node deployable on 2× A100-80GB. Improvements over Llama 3.1 70B include +9.0 on MATH, +7.9 on HumanEval, and +4.6 on IFEval at the same parameter count, signaling a notable efficiency step for the family.

Llama 3.3 70B Instruct is Meta's instruction-tuned, text-only multilingual large language model released on December 6, 2024, built on an optimized dense transformer architecture with 70B parameters (the checkpoint lists 71B) and a 128K context window. It was pretrained on more than 15 trillion tokens with a knowledge cutoff of December 2023, and the model card supports English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai. Meta positioned it as matching Llama 3.1 405B quality at far lower serving cost. The variant uses Grouped-Query Attention for inference scalability, and the released BF16 checkpoints can be loaded in 8-bit or 4-bit via bitsandbytes and transformers. On Meta's own instruction-tuned benchmarks it ties Llama 3.1 70B on MMLU at 86.0, beats the 405B model on IFEval (92.1 vs 88.6) and MATH (77.0 vs 73.8), and is distributed under the custom commercial Llama 3.3 Community License requiring Built with Llama attribution.

Together AI announced on December 6, 2024 that it has partnered with Meta to host Llama-3.3-70B-Instruct, positioning the new 70B release as a significant advance in reasoning, mathematics, and instruction-following with performance comparable to the larger Llama 3.1 405B at a fraction of the cost. The post-training improvements are highlighted as the key driver of those capability gains. To serve the release, Together AI introduced a Turbo serverless endpoint running the model in FP8 for fast inference, alongside forthcoming Dedicated Endpoints offering no rate limits and consistent quality for mission-critical workloads. The announcement notes more than 250,000 developers and enterprises including Zomato, Salesforce, DuckDuckGo, and Washington Post use the Together AI Inference and Fine-Tuning Platform.

Microsoft Foundry lists Meta's Llama-3.3-70B-Instruct as a "Direct from Azure" text-only chat completion model, currently in Preview lifecycle and managed by Microsoft under a single license with enterprise-grade infrastructure. The model targets multilingual dialogue and is described as offering reasoning, math, and instruction-following performance comparable to the much larger Llama 3.1 405B. The Azure catalog reports a 128k context window with an 8,192 output token limit, and benchmarks show MMLU (CoT) at 86.0, MMLU Pro (CoT) at 68.9, IFEval steerability at 92.1, GPQA Diamond reasoning at 50.5, HumanEval code at 88.4, MBPP EvalPlus at 87.6, MATH at 77.0, BFCL v2 tool use at 77.3, and MGSM multilingual at 91.1. Supported input and output modalities are text, with details and pricing available after sign-in.

Meta released Llama 3.3 on December 6, 2024, distributing the foundational model, trained weights, and accompanying Documentation under the Llama 3.3 Community License Agreement. The agreement defines Meta Platforms Ireland Limited or Meta Platforms, Inc. as the rights holder and grants a non-exclusive, worldwide, royalty-free license to use, reproduce, distribute, and create derivative works. Redistribution requires licensees to provide a copy of the Llama 3.3 Community License Agreement to recipients of the Llama Materials or any derivative works, including products or services containing them. Model artifacts and documentation are made available through llama.com, with the Hugging Face repository serving as the canonical access point gated by contact-information consent and the Meta Privacy Policy.

This D-Central model profile covers Meta's Llama 3.3 70B Instruct, naming it explicitly as a single 70-billion-parameter text-only model released by Meta in December 2024 under the Llama family with a Llama 3.3 Community License. The profile confirms the model has a 128,000-token context window, lists primary languages The article frames Llama 3.3 70B Instruct as a post-training refresh of the Llama 3.1 70B backbone rather than a from-scratch retrain, and notes that Meta positions it as matching or exceeding Llama 3.1 405B on most standard benchmarks at roughly a fifth of the active parameters. It also details that the dense transfor