Updated: 2026-07-07

Meta released Llama 3 on April 18, 2024 in two sizes: 8B and 70B, both with Instruct variants for chat. It was trained on more than 15 trillion tokens (7.5x more than Llama 2), with a 128k-vocabulary tokenizer and Grouped Query Attention on both sizes. To me, this was the release where an open model started genuinely competing with closed frontier models on everyday tasks, not just on paper.

Key takeaways

  • 15T training tokens vs 2T for Llama 2: the data scale is the most visible difference in reasoning and instruction following.

  • GQA on 8B and 70B: more efficient inference without sacrificing quality.

  • Llama 3 70B came out ahead of Claude Sonnet, Mistral Medium, and GPT-3.5 in the human evaluation Meta published alongside the launch (1,800 prompts, 12 use-case categories).

  • Llama 3 8B beats Llama 2 13B on almost all benchmarks with half the parameters.

  • The Llama 3 Community License allows commercial use up to 700M MAU at no additional cost.

Key differences from Llama 2

  • 15T training tokens vs 2T: 7.5x more data.

  • Initial 8k context (extended to 128k in Llama 3.1).

  • Improved tokenizer with 128k vocabulary vs 32k: more efficient tokenisation, especially for non-English languages.

  • GQA on both sizes: better quality/inference-cost ratio.

  • Significantly better instruction tuning: SFT + DPO + RLHF, less verbosity, better instruction adherence.

Benchmarks

Benchmark Llama 3 8B Instruct Llama 3 70B Instruct
MMLU (5-shot) 68.4 82.0
GPQA (0-shot) 34.2 39.5
HumanEval (0-shot) 62.2 81.7
GSM-8K (8-shot, CoT) 79.6 93.0
MATH (4-shot, CoT) 30.0 50.4

These are the numbers Meta published in the official Llama 3 model card[1]. One nuance that gets lost in a lot of summaries: Meta never published an exact MMLU or HumanEval table against Claude Sonnet; its open comparison against closed models was via human evaluation, not these academic benchmarks. As the launch blog put it: "Preference rankings by human annotators based on this evaluation set highlight the strong performance of our 70B instruction-following model compared to competing models of comparable size in real-world scenarios" (Meta AI, April 2024[2]). Llama 3 8B, for its part, beats Llama 2 13B on almost everything with half the parameters.

Hardware requirements

Version FP16 INT8 INT4 (GGUF)
Llama 3 8B 16 GB 9 GB 5 GB
Llama 3 70B 140 GB 75 GB 40 GB

8B in Q4 fits in 16 GB Apple Silicon. 70B in Q4 requires an A100 80 GB or two A100 40 GB. For serious production throughput, vLLM is the standard, with tensor parallelism for the 70B; if your workload has a lot of shared prefixes across requests, SGLang is worth a look instead.

Typical deployment

# With Ollama (the simplest option)
ollama run llama3:8b
ollama run llama3:70b

# With vLLM for production
python -m vllm.entrypoints.openai.api_server 
  --model meta-llama/Meta-Llama-3-70B-Instruct 
  --tensor-parallel-size 4 
  --gpu-memory-utilization 0.9

If you don’t want to self-host, Groq serves Llama 3 8B at more than 800 tokens per second according to independent benchmarks (VentureBeat[3]); Together.ai and AWS Bedrock cover the 70B in enterprise mode.

Licence

The Llama 3 Community License allows commercial use up to 700M MAU with a "Built with Meta Llama 3" display requirement. Above that threshold you need to request a special licence from Meta (full terms[4]). For the vast majority of organisations, the licence is permissive enough for production deployments.

Fine-tuning

Llama 3 8B is fine-tunable with manageable resources:

  • QLoRA on a 24 GB A100: workable.

  • LoRA fine-tuning: practical in most enterprise setups.

  • Full fine-tune of the 8B: needs 4x A100 80 GB.

  • Fine-tuning the 70B: H100 cluster, serious budget territory.

For enterprise cases with domain-specific data (legal, medical, technical), a LoRA fine-tune of the 8B on curated data frequently beats a base 70B. See also evaluation frameworks for retrieval for measuring quality after the fine-tune.

Where it excels and where it doesn’t

Strong points:

  • Code generation: HumanEval 62% (8B) and 82% (70B).

  • Maths reasoning: GSM8K 79-93% depending on size.

  • Instruction following: notably better than Llama 2 on adherence and tone.

Relatively weak points:

  • Multilingual: acceptable in Spanish and French, but Mistral and Qwen are still stronger.

  • Long context: 8k base (fixed with Llama 3.1’s 128k).

  • Multimodal: text only in the initial release (fixed with Llama 3.2).

Frequently asked questions

Is Llama 3 free for commercial use?

Yes, up to 700 million monthly active users under the Llama 3 Community License, with the sole requirement of displaying "Built with Meta Llama 3". Above that threshold you need to request a special licence from Meta.

Which Llama 3 size should I pick, 8B or 70B?

The 8B for modest self-hosting (it fits on a 16 GB GPU at Q4) and cases where cost per token dominates. The 70B when reasoning and code quality matter more than cost, since it comes close to closed models in Meta’s human evaluation.

Does Llama 3 support long context?

Not in the initial release: 8k tokens of context. If your use case needs more, wait for Llama 3.1 (128k) or use another model in the meantime.

Conclusion

Llama 3 is a real leap over Llama 2 and sets the open reference standard. The 8B is the default option for modest self-hosting; the 70B competes with closed frontier on most tasks. Combined with a massive ecosystem of fine-tunes, quantised variants, and tooling, it’s the safe choice for teams serious about open LLMs. For extreme multilingual or very long context, Mixtral or Gemini remain preferable; for everything else, Llama 3 is the sensible default.

Spanish version of this article: Llama 3: el nuevo estándar abierto de Meta.

Sources

  1. official Llama 3 model card
  2. Meta AI, April 2024
  3. VentureBeat
  4. full terms