Llama 3.3 70B vs Llama 4 Maverick vs Llama 4 Scout

Best cell per row highlighted. Null means undisclosed — never counted as zero.

Meta
Llama 3.3 70BGA
7.4
AI Panel
Meta
Llama 4 MaverickGA
7.7
AI Panel
Meta
Llama 4 ScoutGA
7.5
AI Panel

Identity & lifecycle

ProviderMetaMetaMeta
Family / tierLlama 3 · LargeLlama 4 · MaverickLlama 4 · Scout
StatusGAGAGA
Released2024-12-052025-04-042025-04-04
Knowledge cutoff2023-122024-082024-08

Architecture & context

Context window128K1M10M
Max output tokens4K8K8K
Input modalitiestexttext, imagetext, image
Reasoning modenonenonenone
Open weightsYesYesYes

Pricing (per Mtok)

Input$0.12$0.2$0.1
Output$0.4$0.85$0.34
Cached input
Batch input
Free tierNoNoNo

Speed

Output speed (tok/s)81.8104.3106.1
Time to first token (s)0.40.660.56
Latency tierfastfastfast

Trust & deployment

Trains on inputsNoNoNo
LicenseLlama-3-CommunityLlama-4-CommunityLlama-4-Community
CloudsBedrock, Vertex AI, Azure AI Foundry, GCP, OCI, IBM watsonxBedrock, Vertex AI, Azure AI Foundry, GCP, OCI, IBM watsonxBedrock, Vertex AI, Azure AI Foundry, GCP, OCI, IBM watsonx
Compliance

AI Panel scoring

Unified score7.47.77.5
Decision Maker88.58.5
Domain Strategist77.57.5
Finance Lead899
Domain Practitioner87.57.5
Power User76.56.5
Skeptic6.55.55.5
Value score8.599.5

Benchmarks

MMLU8685.579.6
MMLU-Pro68.980.574.3
GPQA Diamond50.569.857.2
MATH-5007761.250.3
HumanEval88.485.882
LiveCodeBench43.432.8
Aider Polyglot15.6
MMMU73.469.4
IFEval92.1
LMArena Elo12571271
Artificial Analysis Index141814

Scores link to their sources. Missing cells mean the vendor hasn't disclosed a result — honesty over padding.