What is Llama 4 Maverick?
Llama 4 Maverick is the larger of the two open-weights Llama 4 models Meta released on 5 April 2025, alongside Llama 4 Scout. It is an older line now: Meta replaced Llama with its Muse models in April 2026, has published no Llama since April 2025, and never released Llama 4 Behemoth, the nearly two-trillion-parameter teacher Maverick was distilled from. Meta's newest open model of comparable scale is Muse Spark Open.
Maverick is a mixture-of-experts model with about 400 billion parameters in total and 17 billion active per token, across 128 experts. It reads text and images and writes text and code in 12 supported languages, with a 1-million-token context and an August 2024 knowledge cutoff. Meta ships BF16 and FP8 weights; the FP8 build fits on one eight-GPU H100 host.
The weights are gated on Hugging Face under the Llama 4 Community License. Hosted prices on OpenRouter run from $0.188 input and $0.652 output per million tokens on DigitalOcean to $0.35 and $1.15 on Google Vertex AI. Groq shut it down on 9 March 2026.
How much does Llama 4 Maverick cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | BF16 and FP8, Llama 4 Community License |
| DigitalOcean | $0.188 in / $0.652 out per 1M tokens | 128,000-token context via OpenRouter |
| Parasail | $0.35 in / $1.00 out per 1M tokens | FP8, 524,288-token context |
| Google Vertex AI | $0.35 in / $1.15 out per 1M tokens | 8,192-token output cap via OpenRouter |
Free weights under the Llama 4 Community License (gated). Hosted per 1M tokens: DigitalOcean $0.188 input / $0.652 output; Parasail $0.35 / $1.00; Google Vertex AI $0.35 / $1.15.
Compare this price with other models
Is Llama 4 Maverick worth it?
Maverick's launch is remembered for a benchmark it did not earn. The version that scored 1417 and placed second on LMArena was an unreleased chat-tuned variant, "Llama-4-Maverick-03-26-Experimental"; LMArena said Meta should have made that clearer and changed its rules, and the released model then ranked 32nd, according to TechCrunch. Artificial Analysis put the public model at 49 on its Intelligence Index, later 50, behind DeepSeek V3 0324 at 53.
Meta's own figures are respectable for its time and self-reported: 80.5 on MMLU Pro, 69.8 on GPQA Diamond, 43.4 on LiveCodeBench and 73.4 on MMMU, against Scout's 74.3, 57.2, 32.8 and 69.4. The 1-million-token context is the weak point. On Fiction.LiveBench at 120,000 tokens Maverick scored 28.1%, no better than Llama 3.3 70B, per The Decoder.
Running it yourself is a data-centre job. The BF16 weights are about 803GB and the official FP8 build about 417GB, which Meta says fits on one H100 DGX host of eight 80GB cards; Unsloth's 4-bit GGUF is about 243GB. Only 17 billion parameters are active per token, so it is fast once loaded, but you pay for memory to hold all 400 billion.
The licence is the same as Scout's: a separate Meta licence above 700 million monthly users, "Built with Llama" attribution, "Llama" at the start of any derived model's name, and no multimodal rights for companies based in the European Union. Groq dropped Maverick in March 2026 and pointed customers to gpt-oss-120b, which is Apache 2.0 and runs on a single 80GB GPU. That swap is the likelier path for most teams than a fresh Maverick deployment.
Where it earns its place
A support team at a Southeast Asian marketplace answering in Thai, Indonesian and Vietnamese
Draft replies to customer tickets that include screenshots
All three languages are on Meta's supported list and screenshots go in as images. A million replies a month at 2,000 input and 300 output tokens is about $376 input and $196 output on DigitalOcean, though its 128,000-token context there is shorter.
An infrastructure lead with an eight-GPU H100 node already paid for
Run a larger open model in-house for internal chat and document questions
Serve the official FP8 weights with vLLM across the eight cards, keeping prompts under about 100K tokens where accuracy holds. Compare against gpt-oss-120b on one of the same GPUs before committing all eight.
Before you commit
- Setup
- Accept the licence on Hugging Face and download meta-llama/Llama-4-Maverick-17B-128E-Instruct or the -FP8 build, then serve with vLLM or SGLang across a multi-GPU node; hosted endpoints work immediately through OpenRouter, Parasail, DigitalOcean or Vertex AI.
- Integrations
- OpenAI-compatible APIs from the listed hosts and OpenRouter; tool calling on most endpoints; Google Vertex AI Model Garden; GGUF builds for llama.cpp.
- Data and privacy
- Self-hosted, data stays on your hardware. The Llama 4 Community License adds a 700M-monthly-user threshold, attribution and naming rules, and the acceptable use policy withholds multimodal rights from EU-based individuals and companies, though not their end users.
- Cost at scale
- 1B input and 250M output tokens a month is about $351 on DigitalOcean or $600 on Parasail. Self-hosting: about 803GB in BF16, 417GB in FP8 (one 8x H100 host) and 214 to 243GB at 4-bit.
- Limits
- Advertised 1M-token context but weak past 100K in independent tests; host contexts range from 128,000 to 524,288 tokens; tested with up to 5 images; static model with an August 2024 cutoff.
What’s good
- From $0.188 input and $0.652 output per 1M tokens on DigitalOcean
- 69.8 on GPQA Diamond and 80.5 on MMLU Pro (Meta, self-reported)
- Native image input across 12 supported languages
- Official FP8 weights that fit one eight-GPU H100 host
What’s not
- LMArena ranking at launch came from an unreleased variant
- 28.1% on Fiction.LiveBench at 120K tokens despite a 1M claim
- About 417GB even in FP8; no single-GPU option
- Llama 4 licence terms, including an EU multimodal exclusion; no further Llama releases
Best for
Teams with an existing Maverick deployment or a host contract who want cheap multilingual text and image handling and are outside the EU.
Llama 4 Maverick alternatives
Llama 4 Maverick: common questions
- Is Llama 4 Maverick worth it?
- Maverick was a fair open model in spring 2025, launched under a leaderboard cloud, and nothing Meta has done since gives a reason to start a project on it. We rate it 3.3 out of 5.
- How much does Llama 4 Maverick cost?
- Free weights under the Llama 4 Community License (gated). Hosted per 1M tokens: DigitalOcean $0.188 input / $0.652 output; Parasail $0.35 / $1.00; Google Vertex AI $0.35 / $1.15.
- What are the downsides of Llama 4 Maverick?
- LMArena ranking at launch came from an unreleased variant; 28.1% on Fiction.LiveBench at 120K tokens despite a 1M claim; about 417GB even in FP8; no single-GPU option; llama 4 licence terms, including an EU multimodal exclusion; no further Llama releases.
- Who is Llama 4 Maverick best for?
- Teams with an existing Maverick deployment or a host contract who want cheap multilingual text and image handling and are outside the EU.
- How much does Llama 4 Maverick cost for a team or in production?
- 1B input and 250M output tokens a month is about $351 on DigitalOcean or $600 on Parasail. Self-hosting: about 803GB in BF16, 417GB in FP8 (one 8x H100 host) and 214 to 243GB at 4-bit.
- Is Llama 4 Maverick safe for company data?
- Self-hosted, data stays on your hardware. The Llama 4 Community License adds a 700M-monthly-user threshold, attribution and naming rules, and the acceptable use policy withholds multimodal rights from EU-based individuals and companies, though not their end users.
- What are the best Llama 4 Maverick alternatives?
- The closest alternatives we have reviewed are Muse Spark Open, gpt-oss-120b, DeepSeek-V4.1-Flash, Qwen3.5-397B-A17B and Llama 4 Scout.
Answers checked Nov 9, 2026.