LLM

Llama 3.3 Instruct 70B

by Meta AI
Overall Score
14.5

Meta's leading open-weight model — frontier capability, self-hostable

Released Dec 2024Open Source
Benchmark Scores
Reasoning
8.5
Coding
10.7
Math
7.7
Creative writing
8.4
Instruction following
8.5
Multimodal
7.0
Standard Benchmarks
aa_livecodebenchaa_livecodebench
0.3
Measured 2026-05-13 · source
aa_scicodeaa_scicode
0.3
Measured 2026-05-13 · source
aa_tau2aa_tau2
0.3
Measured 2026-05-13 · source
aa_math_500aa_math_500
0.8
Measured 2026-05-13 · source
aa_aimeaa_aime
0.3
Measured 2026-05-13 · source
aa_aime_25aa_aime_25
0.1
Measured 2026-05-13 · source
aa_ifbenchaa_ifbench
0.5
Measured 2026-05-13 · source
aa_lcraa_lcr
0.1
Measured 2026-05-13 · source
aa_terminalbench_hardaa_terminalbench_hard
0.0
Measured 2026-05-13 · source
aa_mmlu_proaa_mmlu_pro
0.7
Measured 2026-05-13 · source
aa_gpqaaa_gpqa
0.5
Measured 2026-05-13 · source
aa_hleaa_hle
0.0
Measured 2026-05-13 · source
Price
Free for research and most commercial use (see Community License)
Free
Context
128K tokens
Speed
92.41 t/s
580ms TTFT
Compare this model →

Overview

The strongest fully open-weight LLM in regular use. Comparable to GPT-4 class models on most benchmarks while being free to download, fine-tune, and self-host. Foundation for thousands of derivative models.

Strengths

Open weights (Llama Community License) · self-hostable · strong English + multilingual · enormous fine-tune ecosystem · free to use commercially with caveats

Known limitations

Less capable than frontier closed models on hardest reasoning · 128K context smaller than rivals · weaker tool use than purpose-built agentic models

Discussion

0 comments

Sign in to join the conversation.

Be the first to comment.