Llama 3.3 70B
Open weightsby Meta AI·USA·Released
70B-parameter open-weights model matching Llama 3.1 405B quality at a fraction of the cost.
About this model
Llama 3.3 70B (December 2024) is the model that compressed the gap between open and closed weights more than any prior Meta release. At 70B dense parameters it matches the much larger Llama 3.1 405B on most benchmarks — same quality, fraction of the serving cost.
Released under the Llama Community License (commercial use OK with restrictions on companies with >700M MAU), 3.3 became the default open-weights flagship at every major serving provider — Together, Fireworks, Groq, DeepInfra, Cerebras, OctoAI. On Cerebras's wafer-scale silicon it serves at 2000+ tokens/second, faster than any closed-weights model.
For teams that need on-prem deployment, fine-tuning, or just want to avoid vendor lock-in, Llama 3.3 70B has been the default choice for the last year.
Strengths
- •Matches Llama 3.1 405B quality at 1/5 the parameter count
- •Llama Community License — commercial use OK for most companies
- •Served by every major open-weights provider
- •Industry-leading inference speed on specialised hardware (Cerebras 2000+ tok/s)
- •Mature fine-tuning ecosystem
Limitations
- •Llama Community License excludes the largest tech companies (700M MAU clause)
- •Trails Claude 4 / GPT-4.1 on hardest agentic benchmarks
- •128K context — smaller than top closed models
- •No native multimodal — Llama 3.2 family handles vision separately
When to use it
- →On-prem deployments where data residency forbids cloud APIs
- →High-throughput batch inference (Cerebras / Groq workloads)
- →Fine-tuning for vertical specialisation
- →Multi-cloud serving where API portability is required
- →Research applications needing weight access
Architecture & training
70B-parameter dense transformer. Llama 3.3 was released as an instruction-tuned-only update — Meta did not release a separate base model — and represents a substantial post-training improvement over Llama 3.1 70B using techniques including online RLHF and improved data curation. The model card explicitly documents performance vs Llama 3.1 405B across all major benchmarks.
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| MATH | 77.0 | |
| MMLU | 86.0 | |
| HumanEval | 88.4 |