← All models

Meta: Llama 3.3 70B Instruct

Meta

open_source_balanced

Capabilities

ToolsStructured OutputLong Context

Strengths

  • Open weights enable full self-hosting and data privacy
  • Large context window (131K) excellent for long documents
  • Strong multilingual capabilities across tasks
  • Affordable mid-tier pricing for open-weight model

Weaknesses

  • Slow inference speed (96 tok/s) limits real-time applications
  • Weak coding and reasoning performance (0.3, 0.27)
  • No vision capability for multimodal tasks
  • Limited mathematical reasoning (0.39 fitness)

Pricing

Input / 1M tokens

$0.09999999999999999

Output / 1M tokens

$0.32

Context window

131.072k

Transparency

Open weights

10.0 / 10

Open training data

0.0 / 10

Open methodology

7.5 / 10

Licence openness

6.0 / 10

Provider disclosure

7.0 / 10

FMTI company score

N/A

Composite:7.0 / 10

Llama 3.3 ships with publicly downloadable weights under Meta's Llama License. Meta publishes detailed model cards and training methodology. Training data composition not fully disclosed. Strong open-weight positioning enables self-hosting and full data privacy control.

Sustainability

Inference energy

N/A

Training footprint

N/A

Provider infrastructure

5.5 / 10

Composite:5.0 / 10

70B parameter count suggests moderate inference energy relative to flagship models. Meta's infrastructure sustainability metrics not widely published. Self-hosting capability reduces reliance on provider infrastructure but does not offset unknown training footprint.

Task Fitness

question answering
52%
code
30%
mathematics
39%
business communication
62%
vision
0%
analysis
44%
extraction
56%
long-form generation
65%
research
54%
reasoning
27%
summarisation
75%
translation
62%
conversation
68%