Llama 4 Maverick

metameta/llama-4-maverick

Llama 4 Maverick (400B total, 17B active, 128 experts MoE) offers industry-leading performance in image and text understanding with support for 12 languages. Great for precise image understanding and creative writing. Our product workhorse model for general assistant and chat use cases.

Meta AI blog — Llama 4 launch ↗Hugging Face model card ↗Llama 4 models overview ↗Llama 4 Community License ↗

Llama 4 Maverick, released April 5 2025, is Meta's flagship open-weight multimodal language model. It uses a 128-expert mixture-of-experts (MoE) architecture with 400 billion total parameters and only 17 billion active parameters per forward pass, delivering frontier-class text and image reasoning at a fraction of the inference cost of dense models at similar capability levels. The model supports a 1 million token context window, 12 languages, and up to 5 simultaneous image inputs.

From a TheRouter routing perspective, Maverick is the highest-capability Meta model available at the $0.30/$1.25 per million token price point — a 400× parameter advantage over comparable dense models at similar cost, achieved through MoE sparsity. It excels at precise image understanding, multilingual assistant tasks, and long-context document work. For tasks where the full 1M context is unnecessary, its sibling Llama 4 Scout (17Bx16E, 10M context) offers lower per-token cost at lighter workloads. Maverick's open weights and Llama 4 Community License make it suitable for both commercial deployment and self-hosted inference pipelines.

Best for

• Multimodal reasoning tasks combining text and images, such as diagram analysis, visual QA, and document parsing with embedded figures
• Long-context document analysis up to 1M tokens — large codebase review, contract synthesis, multi-report summarization
• Multilingual assistant and translation workloads across the 12 natively supported languages including Arabic, Hindi, and Indonesian
• Self-hosted inference pipelines where open weights, commercial licensing freedom, and cost-efficient MoE serving matter
• Creative writing and general assistant chat where high throughput and low latency per active parameter matter more than state-of-the-art reasoning depth

Reach for something else if

• Hard reasoning chains and math olympiad-level problems — route to deepseek/deepseek-r1 or anthropic/claude-opus-4.7 for those
• Production agentic coding with tool-use loops requiring top SWE-bench reliability — use anthropic/claude-sonnet-4.6 or openai/gpt-5.2-codex for that
• Ultra-long context (beyond 1M tokens) — route to meta/llama-4-scout which supports a 10M token window

How TheRouter serves this differently from the vendor

As the vendor operates it

Meta releases Llama 4 Maverick as open weights for self-hosting and also references it as a Meta AI model. Operators can run the BF16 or FP8 checkpoints directly under the Llama 4 Community License.

On TheRouter

TheRouter serves the model as an OpenAI-compatible chat and vision endpoint at api.therouter.ai/v1. It does not expose Meta's raw checkpoints, self-hosting controls, training-time safety stack, or any Meta AI consumer-product surface.

github.com/meta-llama ↗

Context Length

Max Output

16K

Input Priceper 1M tokens

$0.2592/ 1M tokens

Output Priceper 1M tokens

$1.05/ 1M tokens

Modalities

textimage→text

Capabilities

Vision

Pricing Breakdown

Type	Rate
Input	$0.2592 / 1M tokens
Output	$1.05 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatstop

Specifications

Release date	April 5, 2025huggingface.co ↗	verified
Architecture	Mixture-of-Experts (MoE) — 128 experts, alternating dense + MoE layersai.meta.com ↗	verified
Total parameters	400Bhuggingface.co ↗	verified
Active parameters per forward pass	17Bai.meta.com ↗	verified
Context window	1,000,000 tokens (1M)huggingface.co ↗	verified
Training token count	~22 trillion tokenshuggingface.co ↗	verified
Knowledge cutoff	August 2024huggingface.co ↗	verified
Supported languages	12 natively (Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai, Vietnamese)huggingface.co ↗	verified
License	Llama 4 Community License (custom commercial license; Meta products/services usage permitted)github.com/meta-llama ↗	verified
Multimodal input support	Text + image (up to 5 images per request)huggingface.co ↗	verified

Benchmarks

Benchmark	Score	Source
MMLU Pre-trained model evaluation; 5-shot macro_avg/acc_char reported by Meta for Llama 4 Maverick.	85.5%%	github.com/meta-llama ↗
MMLU Pro Instruction-tuned model evaluation; 0-shot macro_avg/acc reported by Meta.	80.5%%	github.com/meta-llama ↗
GPQA Diamond Instruction-tuned model evaluation; 0-shot accuracy reported by Meta.	69.8%%	github.com/meta-llama ↗
MATH Pre-trained model evaluation; 4-shot em_maj1@1 reported by Meta.	61.2%%	github.com/meta-llama ↗
MMMU Instruction-tuned image reasoning evaluation; 0-shot accuracy reported by Meta.	73.4%%	github.com/meta-llama ↗
MathVista Instruction-tuned image reasoning evaluation; 0-shot accuracy reported by Meta.	73.7%%	github.com/meta-llama ↗
ChartQA Instruction-tuned image understanding evaluation; 0-shot relaxed_accuracy reported by Meta.	90.0%%	github.com/meta-llama ↗
MBPP Pre-trained code evaluation; 3-shot pass@1 reported by Meta.	77.6%%	github.com/meta-llama ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL

curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "meta/llama-4-maverick",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference →

Chat completion

Standard OpenAI-compatible chat endpoint. Switch only the baseURL and model name — no SDK changes required.

cURL

curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta/llama-4-maverick",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Similar models

Cross-provider sibling models

google/gemini-2.5-flash-lite

Google's MoE-based efficient multimodal model; comparable price tier, different provider trade-offs

deepseek/deepseek-v3.1

Deepseek's dense-text frontier model; stronger coding but no native image input

News & changes

2025-04-05

Meta releases Llama 4 Maverick and Scout — open multimodal MoE herd

Meta launched the Llama 4 family on April 5, 2025, introducing two open-weight multimodal models: Maverick (17Bx128E, 400B total, 1M context) and Scout (17Bx16E, 109B total, 10M context). Both use early-fusion MoE architectures enabling native image + text reasoning, trained on ~22T and ~40T tokens respectively, with a knowledge cutoff of August 2024. The release established Meta as a competitive open-weight multimodal player against proprietary frontier models.

re-authored by TheRouterai.meta.com ↗

Frequently asked

What is the difference between Llama 4 Maverick and Llama 4 Scout?

Both share the same 17B active parameter count, but Maverick has 128 experts (400B total) vs. Scout's 16 experts (109B total), giving Maverick significantly stronger performance on complex reasoning, coding, and multimodal tasks. Scout compensates with a 10× longer context window (10M vs. 1M tokens) and lower inference cost, making it preferable for ultra-long-document tasks where context range matters more than per-token reasoning quality.

huggingface.co ↗

Can I self-host Llama 4 Maverick?

Yes — the weights are publicly available on Hugging Face under the Llama 4 Community License, which permits commercial use. With 400B total parameters and 17B active per forward pass, Maverick requires distributed inference infrastructure (typically multi-GPU serving via frameworks like vLLM or TensorRT-LLM). For teams that prefer managed API access without infrastructure overhead, TheRouter provides on-demand access at $0.30/$1.25 per million tokens with no GPUs to operate.