Meta's open-weight MoE model pairing text-and-image input, a 1M-token window, and fast, low-latency inference for high-volume assistants.
Modality:
Text
Image
model ID
meta-llama/llama-4-maverick
Output Speed *
68.64
tok/s
Intelligence Index *
10
/ 100
Context Window *
1048576
tokens
Input price
0.9
Anytoken
Output price
3.6
Anytoken
Llama 4 Maverick: Fast, Open-Weight Multimodal Inference at Scale
Llama 4 Maverick is Meta's larger Llama 4 model, a natively multimodal mixture-of-experts system with 17B active parameters drawn from 400B total across 128 experts. It sits above Llama 4 Scout in capability while sharing the same active-parameter count, so inference stays fast. Maverick accepts text and image input, returns text, and supports a 1M-token context window. Its strongest fit is high-throughput, cost-sensitive production workloads—multilingual assistants, document and image understanding, and streaming chat—where fast output and low first-token latency matter more than frontier reasoning.
Start calling Llama 4 Maverick through the AnyAPI.ai API
Performance
Where Maverick Earns Its Place: Speed and Multimodal Throughput
Maverick's defining strength is fast, low-latency multimodal serving rather than frontier intelligence. Independent measurements place median output around 100 tokens per second—well above the roughly 60 t/s median for comparable open-weight non-reasoning models—with sub-second time to first token on optimized FP8 providers. Because the MoE architecture activates only 17B parameters per token, throughput stays high despite the 400B total. For production, this means Maverick suits streaming chat, high-volume assistants, and image-plus-text pipelines where responsiveness and cost efficiency dominate, and where you don't require top-tier reasoning or agentic accuracy.
Benchmarks
Llama 4 Maverick Benchmarks: Fast and Efficient, Not Frontier
Independent testing tells a consistent story. Artificial Analysis scores Maverick below the median on its Intelligence Index while rating it notably fast and reasonably priced among open-weight non-reasoning peers. Meta's reported model-card figures show strong multimodal results (MMMU 73.4, MathVista 73.7) and competitive MMLU Pro, but agentic and novel-reasoning tests are weak—Maverick scored near zero on ARC-AGI-2 in independent runs, and coding percentiles sit low. Note that Meta's high LMArena ELO came from an experimental, style-tuned variant; the public model ranked far lower. Evaluate on your own tasks.
Output Speed
*
68.64
tok/s
Intelligence Index
*
10
/ 100
MMLU *
Broad world knowledge and problem-solving
81
%
GPQA *
PhD-level scientific reasoning across physics, biology, chemistry.
67
%
HLE *
Adherence to multi-step structured instructions.
5
%
LiveCodeBench *
Tool-calling reliability in long agentic loops.
40
%
Technical Specifications
What the model supports
Maverick accepts text and image input and returns text only; image understanding is English-only, even though text spans 12 fine-tuned languages. The 1M-token context window is large, but maximum output is capped at 8,192 tokens, so it fits long-context reading far better than long-form generation. It is a non-reasoning model with no extended-thinking mode. Tool calling and JSON-schema structured outputs are supported. Note the knowledge cutoff of August 2024 and the license restriction prohibiting EU-domiciled use, which affects deployment planning.
Verified Specifications —
Llama 4 Maverick
*
Input modalities
Text
Image
output modalities
Text
Context window
1048576
tokens
Maximum output tokens
16384
Reasoning
No
Knowledge cutoff
April 2025
Pricing (standard)
0.9
AnyTokens in
/
3.6
AnyTokens out
Limitations & Trade-offs
1
Below-average intelligence for demanding tasks. Independent evaluation places Maverick under the median on the Artificial Analysis Intelligence Index and low on coding and math percentiles. For complex reasoning, agentic planning, or hard coding, frontier or reasoning-capable models will outperform it. Maverick is a fast, general-purpose workhorse, not a reasoning specialist—route difficult branches elsewhere.
2
No reasoning mode and weak agentic scores. Maverick has no extended-thinking configuration, and independent agentic benchmarks (including near-zero ARC-AGI-2 results) show it is a poor fit for autonomous multi-step agents. If your workload depends on reliable tool-chaining, long planning horizons, or self-correction, a dedicated reasoning or agentic model is preferable despite Maverick's speed and cost advantages.
3
Short maximum output despite a huge context window. The 1M-token context is paired with an 8,192-token output cap, so Maverick reads far more than it can write in one turn. Long-form generation—large reports, extensive code files, book-length drafts—requires chunking or a different model. Treat the large window as an input/retrieval advantage, not a long-output capability.
4
Deployment and licensing constraints. Self-hosting Maverick realistically needs a full H100 DGX-class host (FP8) rather than a single consumer GPU, and the Llama 4 Community License restricts EU-domiciled use and imposes a 700M-monthly-active-user threshold. Teams in the EU, or those without enterprise GPU access, should plan around hosted APIs and confirm license terms before production.
Best-Fit Workloads
Where this model earns its place
01
High-volume multilingual chat assistants
Maverick's fast output (~100 t/s median) and low first-token latency make it well suited to streaming, interactive assistants at scale. With 12 fine-tuned languages and strong steerability via system prompts, it handles multilingual customer-facing chat cost-efficiently. Keep responses within the 8,192-token cap and route genuinely hard reasoning to a stronger model.
02
Image and document understanding
Native text-and-image input supports visual Q&A, captioning, chart and document analysis, with strong reported multimodal benchmarks (MMMU 73.4, MathVista 73.7). Combined with the large context window, Maverick fits pipelines that ingest images alongside long text. Note image understanding is English-only, which constrains non-English visual workloads.
03
Long-context retrieval and RAG
The ~1M-token window lets Maverick ingest large documents, transcripts, or retrieved passages in a single request, while fast inference keeps RAG responses snappy. It excels at reading and synthesizing long inputs rather than producing long outputs, so pair it with retrieval and keep generations concise. Independent testing suggests validating long-context accuracy on your own data.
04
Latency-sensitive prefill and first-turn responses
With the lowest measured time to first token in several independent comparisons, Maverick is a strong choice for autocomplete, predictive typing, and warm-up turns where perceived responsiveness dominates. In multi-model routes it can take the first turn or easy branches while smarter, slower models handle difficult reasoning steps.
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04