Llama 4 Maverick
Open SourceMeta•Released on 2025-04-05
Meta's flagship open-source multimodal model. 17B active parameters with 400B total (128 expert MoE). 1M context window, natively multimodal with early fusion. Extremely cost-effective at $0.15/$0.60 per M tokens. Supports 12 languages.
80
Overall Score
Voice of the community
sample 892“Llama 4 Maverick quickly secured the number-two spot on LMArena with an ELO score of 1417.”
“Llama 4 Maverick achieved only a 70% accuracy score on coding benchmarks, underperforming against coding-centric models.”
“The price-to-performance ratio is insane. For simple tasks and high-volume workloads, nothing beats Llama 4.”
Core Specs
1049K
Context Window
16K
Max Output
ReasoningOpen Sourcetextimage
Scenario Scores
Pros & Cons
Sentiment65% +20% ·15% −
Pros
- +Extremely affordable ($0.15/$0.60)
- +1M context window
- +Native multimodal (text + image)
- +Open source (Llama 4 Community License)
- +High throughput MoE architecture
Cons
- −Coding performance below Claude/GPT
- −Benchmark gaming controversy
- −16K max output limit
- −Knowledge cutoff August 2024
Pricing
Input (per 1M tokens)$0.15
Output (per 1M tokens)$0.60
Free trial available
Updated on 2026-03-07
Get Started with Llama 4
2Or self-host (open source)
💡 Open source - can be self-hosted or used via API providers.
Benchmarks
mmmu73.4%
codingAccuracy70%
lmArenaElo1417%
Tools Supporting This Model
Reliability
Incidents (30d)0
开源模型，可靠性取决于托管方