Table of Contents
At a glance
- What it is: Meta’s open-weight, natively multimodal mixture-of-experts model, released April 5, 2025. It has 17B active parameters, about 400B total, and 128 routed experts plus a shared expert.
- Inputs and outputs: Text and up to five images in; text and code out. Twelve languages are officially supported, and the knowledge cutoff is August 2024.
- Context window: 1 million tokens on the model card. Individual hosts often set lower limits.
- License: The Llama 4 Community License. It is open-weight, not open source, with a 700 million user cap, attribution rules, and a multimodal carve-out for entities based in the European Union.
- Reputation: Its launch was shadowed by a leaderboard dispute. Meta’s own former chief AI scientist later said results were fudged “a little bit,” according to news reports.
- Status today: Still available, but Amazon Bedrock lists it as a legacy model with an end-of-life date of March 30, 2027, and Meta’s newer releases carry the Muse name.
Few open-weight models arrived with as much noise as LLaMA 4 Maverick. The company presented it at launch as the leading model among its multimodal peers. Within days, a leaderboard dispute turned the conversation from capability to trust. A year and a half later, the dust has settled enough to look at the model on its own terms.
This guide covers what Llama 4 Maverick is, what its specifications imply, what its published benchmarks do and do not show, how to run it, what the license requires, and where it fits now that Meta has moved on to a different model line. It is written for developers, technical leads, and AI-curious readers who want a fair, evidence-based picture.
A note on method. Everything below comes from Meta’s launch post, the official model card, cloud provider documentation, license text, and clearly labeled news and community sources. Nothing here describes personal experience running the model. Where a figure comes from a third party, the text says so. AI products change quickly, so confirm prices, availability, and lifecycle dates with each provider before you commit.
What LLaMA 4 Maverick Is
Llama 4 Maverick is the bigger of the two models Meta introduced on April 5, 2025, with Llama 4 Scout being the other release. Meta described Maverick as its “product workhorse model for general assistant and chat use cases,” suited to precise image understanding and creative writing. The company also previewed a third, much larger model, Llama 4 Behemoth, which it described as a teacher model still in training.
Maverick was Meta’s first model built with a mixture-of-experts design, and its first open-weight model that takes text and images through one backbone. The official model is named Llama-4-Maverick-17B-128E-Instruct on Hugging Face, where “17B” is the active parameter count and “128E” is the number of experts.
Where It Sits in the Llama 4 Family
The three models share a design idea but target different jobs. The following comparison draws on Meta’s launch announcement and model card:
| Model | Active parameters | Total parameters | Experts | Context | Status |
| Llama 4 Scout | 17B | About 109B | 16 | 10M tokens | Released |
| Llama 4 Maverick | 17B | About 400B | 128 | 1M tokens | Released |
| Llama 4 Behemoth | 288B | Nearly 2T | 16 | Not specified | Previewed, not released |
Scout is the small, long-context model. Meta said it fits on a single H100 GPU with Int4 quantization. Maverick is the quality-focused model, which Meta said fits on a single H100 DGX host. Behemoth acted as a teacher: Meta says Maverick was codistilled from it.
A Short Timeline
The public record gives a clear sequence. Meta released Scout and Maverick on April 5, 2025. Within days, the leaderboard dispute described later in this guide broke out. Meta said in the launch post that it would share more at LlamaCon on April 29. Google Cloud’s documentation lists the Maverick managed API as generally available with an April 29, 2025 release date. News coverage reports Behemoth was postponed in May 2025, and as of this writing it has not shipped as open weights. Meta’s later model line, covered near the end, moved in a different direction.
What the Specifications Tell Us
The model card is unusually specific, so it is worth reading closely. Here are the headline facts:
| Specification | Llama 4 Maverick |
| Architecture | Auto-regressive, mixture-of-experts, early-fusion multimodal |
| Active / total parameters | 17B / about 400B |
| Experts | 128 routed experts plus 1 shared expert |
| Layers | Alternating dense and MoE layers |
| Input | Multilingual text and image |
| Output | Multilingual text and code |
| Context length | 1M tokens |
| Pre-training tokens | About 22 trillion |
| Knowledge cutoff | August 2024 |
| Supported languages | Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai, Vietnamese |
| Release date | April 5, 2025 |
| Weights | BF16 and FP8 |
Mixture of Experts in Plain Terms
In a dense model, each token passes through the model’s entire set of parameters. A mixture-of-experts model keeps a large pool of specialist sub-networks and sends each token through only a few of them. Meta’s explanation is that each token goes to the shared expert and to one of the 128 routed experts, so while all parameters sit in memory, only a subset does work for any one token.
Two practical consequences follow. Compute and latency per token look like those of a roughly 17B-parameter model, which is why Meta pitched Maverick as offering high quality at a lower price than Llama 3.3 70B. But memory is a different story. All 400B parameters must be loaded, so memory needs track total parameters, not active ones. People often miss this when they read “17B” and imagine a laptop-sized model.
Native Multimodality
Meta describes Llama 4 as natively multimodal, using early fusion so text and vision tokens share one backbone and can be pre-trained together on text, image, and video data. The vision encoder is based on MetaCLIP but was trained separately alongside a frozen Llama model to fit the language model better.
The model card gives the practical limits. Meta says the models were pre-trained on up to 48 images, and the card says image understanding was evaluated for up to 5 input images. It adds that developers going beyond that are responsible for extra evaluation and mitigation. Output is text only, so Maverick reads images but does not generate them. Meta’s documentation page also notes that image understanding is English-only.
The Context Window Question
The model card lists a 1 million token context length for Maverick. The reality depends on where you run it. Amazon Bedrock lists a 1M-token context window with an 8K maximum output. One routing marketplace listing, captured a long time ago, showed a 524K window for the same model. A Groq documentation page mentions a 128K window. These differences come from provider configuration and memory limits, not from different models.
The lesson is simple. Read the provider’s limits before you design around a long context. For applications that need to ground model responses in external knowledge rather than rely on the model’s context window alone, see our guide to vector search, RAG, and agentic AI. Also remember that Meta trained Scout, not Maverick, for the record 10M window.
Training and Post-Training
The card says Maverick consumed 2.38 million GPU hours on H100 hardware, with an estimated 645 tons of location-based emissions, and that the pre-training mix was publicly available data, licensed data, and information from Meta’s products and services, including public Instagram and Facebook posts and interactions with Meta AI.
Meta’s launch post describes the post-training pipeline in some detail: lightweight supervised fine-tuning, then online reinforcement learning, then lightweight direct preference optimization. Meta reports that it discarded over half of the data labeled as easy by using Llama models to evaluate it. The reason was that excessive fine-tuning constraints could reduce the model’s ability to explore during reinforcement learning. Treat these as Meta’s own claims about its recipe, not independently verified findings.
Meta also says pre-training covered 200 languages, with more than 100 of them above one billion tokens each, and that this is ten times the multilingual tokens of Llama 3. Only 12 languages are officially supported, though the card allows developers to fine-tune for others under the license.
Published Benchmark Comparison
Meta publishes a table of instruction-tuned results on the model card. Here is a selection, with the predecessors Meta chose to compare against:
| Benchmark | Llama 3.3 70B | Llama 3.1 405B | Llama 4 Scout | Llama 4 Maverick |
| MMLU Pro | 68.9 | 73.4 | 74.3 | 80.5 |
| GPQA Diamond | 50.5 | 49.0 | 57.2 | 69.8 |
| LiveCodeBench | 33.3 | 27.7 | 32.8 | 43.4 |
| MGSM | 91.1 | 91.6 | 90.6 | 92.3 |
| MMMU | No image input | No image input | 69.4 | 73.4 |
| MathVista | No image input | No image input | 70.7 | 73.7 |
| ChartQA | No image input | No image input | 88.8 | 90.0 |
| DocVQA (test) | No image input | No image input | 94.4 | 94.4 |
The card also reports long-context translation scores on a task called MTOB, where Maverick reaches 54.0 and 46.4 on the half-book setting across the two directions, compared with Scout’s 42.2 and 36.6.
What the Numbers Support
Three conclusions are reasonable. First, within the Llama family, Maverick clearly improved on Llama 3.3 70B and on the older, much larger Llama 3.1 405B on reasoning and coding measures. Second, its image understanding scores were competitive for the period. Third, the one place where Scout matches it is DocVQA, which suggests document reading did not need the larger model.
The launch post went further, claiming Maverick beat GPT-4o and Gemini 2.0 Flash across a broad range of widely reported benchmarks and matched DeepSeek V3 on reasoning and coding with less than half the active parameters. Those comparisons were against models from early 2025. They say little about how Maverick stands against current systems.
Developers exploring more complex, tool-using coding systems can also read our Claude Agent SDK developer guide for a practical look at building autonomous AI agents.
Independent Signals
The model card lists only Meta’s own numbers, so outside measurements matter. A news report on Meta’s later Muse Spark launch cited the Artificial Analysis Intelligence Index v4.0, noting that Llama 4 Maverick debuted with an index score of 18, while Muse Spark scored 52. The Hugging Face page also displays community-submitted evaluation entries, including 80.5 on MMLU-Pro (matching Meta’s figure), about 5.2 on SWE-Bench Pro, and about 5.7 on Humanity’s Last Exam. Those last two reflect hard agentic and expert-level tasks where the model was never positioned to lead.
The pattern is consistent. On general knowledge and multimodal reading, Maverick looks respectable. On hard coding agents and expert-level reasoning, it sits well behind current frontier systems. That is a fair reading of the record, and it matches Meta’s own positioning of Maverick as a general assistant model, not a reasoning specialist.
These distinctions become especially important when evaluating autonomous systems; our guide to AI agents vs. chatbots explains how their architectures, tool access, and use cases differ.
The LMArena Episode
No honest review of this model can skip the leaderboard dispute, because it shaped how people read every other number.
What Happened
Meta’s launch post said an experimental chat version of Maverick scored an Elo of 1417 on LMArena. Reporters and researchers soon noticed that the entry, labeled Llama-4-Maverick-03-26-Experimental, was not the downloadable model. LMArena posted that Meta’s interpretation of its policy did not match what it expects from model providers, and said Meta should have made clearer that the entry was customized to optimize for human preference. LMArena said it would update its leaderboard policies.
When the unmodified release, Llama-4-Maverick-17B-128E-Instruct, was added on April 11, 2025, TechCrunch reported it ranked below several months-old models from other labs, and secondary sources place it around 32nd. Meta’s spokesperson said the experimental model was a chat-optimized version the company experimented with, and a Meta executive said it was simply not true that the company trained on test sets.
What Was Said Later
In January 2026, the Financial Times reported an interview with Yann LeCun, who had left Meta as chief AI scientist. News coverage of that interview quotes him saying the results were fudged a little bit and that different models were used for different benchmarks. That account comes from a single interview as relayed by outlets such as Slashdot. Meta’s statements at the time of the launch were more limited, so treat the later account as reported, not as a documented audit.
How to Read It
The fair reading has two parts. The numbers on the model card are Meta’s evaluations of the released weights, and the community-submitted MMLU-Pro entry lines up with them. The LMArena figure, though, belonged to a different variant. The takeaway for buyers is practical: ask for the exact model identifier behind any score, and prefer measurements you can reproduce on your own data.
Hardware and Deployment
How Much Memory the Model Needs
Meta’s position is that Maverick, released in BF16 and FP8 quantized weights, fits on a single H100 DGX host with the FP8 version, and it provides code for on-the-fly Int4 quantization to reduce memory further. A DGX H100 system contains eight GPUs with 80 GB of memory each. This means a “single host” refers to a powerful server, not a typical workstation.
Community calculators give rougher numbers for weights alone. One VRAM calculator estimates about 800 GB at FP16, about 400 GB at INT8, and about 200 GB at INT4, and says the KV cache for long contexts adds a great deal on top. Treat these as planning estimates, not guarantees. The structural point holds either way: memory scales with the 400B total, so a model that is cheap per token can still demand expensive hardware to host.
Self-Hosting Options
The Hugging Face model card says Transformers v4.51.0 or later is required and shows loading through the Llama4ForConditionalGeneration class. It also lists vLLM and SGLang commands for serving an OpenAI-compatible endpoint, and notes that quantized community versions exist for tools such as llama.cpp, Ollama, and LM Studio. Access requires accepting the Llama license on the model page first.
Managed Cloud Access
Most teams reach Maverick through a cloud provider. Here is how the main routes compare, based on provider documentation:
| Provider | Model ID or name | Notes from documentation |
| Amazon Bedrock | us.meta.llama4-maverick-17b-instruct-v1:0 | 1M-token context, 8K max output, knowledge cutoff August 2024; marked Legacy with an end-of-life date of March 30, 2027 |
| Google Cloud | llama-4-maverick-17b-128e-instruct-maas | Managed API, generally available; supports text, code, and image input; function calling and batch predictions listed |
| Oracle Cloud | meta.llama-4-maverick-17b-128e-instruct-fp8 | Hosted on dedicated clusters; not available for fine-tuning there |
| Groq | meta-llama/llama-4-maverick-17b-128e-instruct | Documentation lists up to 5 image inputs |
The Bedrock lifecycle note deserves attention. A legacy label with a published end-of-life date means you should plan a migration path before building anything long-lived on that platform.
What It Costs
Prices vary by host and change often, so treat any figure as a snapshot. One routing marketplace listing showed roughly 19 cents per million input tokens and 65 cents per million output tokens, while another aggregator listed starting prices around 24 and 77 cents. Those listings were captured at different times. Check the current price on the provider you intend to use, and compare it with the cost of newer models that may now outperform Maverick at a similar price.
License and Compliance
Meta calls the Llama 4 Community License a custom commercial license. It is generous in some ways and restrictive in others, so read it before you build a product.
What the License Allows
Section 1 grants a non-exclusive, worldwide, royalty-free license to use, reproduce, distribute, copy, modify, and create derivative works of the Llama materials. The model card says the license also covers using Llama outputs to improve other models, including synthetic data generation and distillation. Under Section 5, you own derivative works you create, as between you and Meta.
What It Requires
The obligations are specific:
| Requirement | What it says |
| Attribution | Display “Built with Llama” on a related website, interface, blog post, about page, or documentation when you distribute Llama materials or a product containing them |
| Naming | If you use the materials or their outputs to create or improve an AI model that you distribute, begin that model’s name with “Llama” |
| License copy and notice | Provide a copy of the agreement and keep the attribution notice in a Notice file |
| Acceptable Use Policy | Comply with Meta’s policy, which is incorporated by reference |
| 700 million user threshold | If your products had more than 700 million monthly active users in the month before the release date, you must request a license that Meta may grant at its sole discretion |
| Litigation clause | If you sue Meta alleging the materials or outputs infringe your IP, your licenses terminate |
The European Union Carve-Out
The Llama 4 Acceptable Use Policy says that for the multimodal models, the Section 1(a) rights are not granted to individuals domiciled in, or companies with a principal place of business in, the European Union. A Swiss compliance guide notes that a non-EU company’s EU-based employees may still work with the models within the scope of that company’s license, and reporting on the license says end users of products built on the models are excepted. The effect is a gap for EU-based developers who want the vision features, while text-only paths are described by commentators as usable. This area is legally sensitive, so involve counsel instead of relying on a summary.
Open Weight Is Not Open Source
The Open Source Initiative’s position, as summarized by several commentators, is that the Llama license does not meet the Open Source Definition because of the user threshold and the usage policy. The accurate label is open-weight with a custom license. That distinction matters for procurement teams that have policies about open source software.
Safety, Safeguards, and Limitations
Meta’s launch post describes mitigations at each layer: data filtering in pre-training, policy-focused post-training, and system-level tools. It includes Llama Guard to assess incoming and outgoing content, Prompt Guard to identify jailbreak attempts and prompt injections, and CyberSecEval to evaluate security. Code Shield is also listed in the model card. Meta also describes an automated red-teaming approach it calls Generative Offensive Agent Testing, which simulates multi-turn adversarial conversations.
Organizations evaluating these safeguards as part of a wider AI risk program can also explore our guide to AI governance tools, which compares governance platforms, runtime guardrails, and monitoring approaches.
On cyber risk, the model card concludes that Llama 4 models do not introduce risk plausibly enabling catastrophic cyber outcomes. On bias, Meta reports that Llama 4 refuses less on debated political and social topics, from 7% for Llama 3.3 to below 2%, and says its rate of strongly leaning responses is comparable to Grok’s. Those are Meta’s own measurements, so apply your own evaluation before relying on them for sensitive audiences.
The card is also clear on limits. Maverick is a static model trained on an offline dataset with an August 2024 cutoff. It supports twelve languages officially. It may give inaccurate or objectionable responses, and Meta tells developers to perform safety evaluation and tuning for their specific application. A model with this cutoff will not know about later events, libraries, or product versions unless you supply them in context.
Feature Comparison: How These Products Compare
Here is how Maverick compares with its siblings and its Llama 3 predecessors, using Meta’s published specifications:
| Feature | Llama 3.1 405B | Llama 3.3 70B | Llama 4 Scout | Llama 4 Maverick |
| Architecture | Dense | Dense | Mixture of experts | Mixture of experts |
| Active parameters | 405B | 70B | 17B | 17B |
| Total parameters | 405B | 70B | About 109B | About 400B |
| Image input | No | No | Yes | Yes |
| Context window | 128K | 128K | 10M | 1M |
| Pre-training tokens | Not listed here | Not listed here | About 40T | About 22T |
| Best fit | Large dense text model | Text assistant on modest hardware | Long documents, single-GPU hosting | General multimodal assistant |
The interesting row is active parameters. Maverick’s per-token compute resembles a mid-size model, yet its total size resembles Llama 3.1 405B. That mix is the heart of the tradeoff: better quality per unit of compute, paid for with memory.
Where Meta Went Next
Meta’s newer announcements changed the context around Llama 4. According to VentureBeat’s coverage, Meta launched Muse Spark, a proprietary model, as its first release from Meta Superintelligence Labs. When asked about Llama’s future, a Meta spokesperson said its current Llama models would continue to be available as open source, which does not address future Llama versions. Secondary coverage dates Muse Spark to April 8, 2026, and reports a later open-weight release called Muse Glimmer, a 30B model under the Apache 2.0 license, in August 2026. These dates come from third-party guides, so confirm them on Meta’s own pages.
Two points stand out. Llama 4 Behemoth, previewed as a teacher model, had not been released as of the most recent guides. And Maverick, once Meta’s open-weight flagship, now reads as a previous-generation model. That does not make it useless. It means you should compare it with newer open-weight options before you start a new project.
For a broader comparison of competing open-weight model families, see our guide to Qwen 3 vs. Llama 4, which examines their architectures, context windows, and relative strengths.
Our Analysis: Who Should Still Consider Maverick
Based on the specifications and published results, here is a reasoned fit guide. It is a reading of the record, not a ranking:
| Situation | Fit | Why |
| Existing system already built on Maverick | Reasonable to keep, with a migration plan | Lifecycle dates on managed platforms are published |
| General chat assistant with image understanding | Plausible | Strong chart and document scores for its period |
| Long-document retrieval over millions of tokens | Check Scout or other options | Scout has the 10M context; Maverick is listed at 1M |
| Hard coding agents and expert reasoning | Weak fit | Community-listed results on hard agentic and expert tasks are low |
| Teams based in the European Union needing vision | Needs legal review | The multimodal rights carve-out applies |
| Teams needing a permissive, open-source license | Poor fit | Custom license with a user threshold |
| Teams with limited GPU memory | Poor fit unless using a managed API | 400B total parameters must be loaded |
Practical Tips Before You Build
A few habits follow from the documentation.
- Confirm the context and output limits on your provider. The model card says 1M, but hosts differ, and Bedrock caps output at 8K tokens.
- Keep images to five or fewer per request. That is the evaluated range on the model card.
- Start from Meta’s suggested system prompt. The card says a system prompt can reduce false refusals and templated language, and it includes a template you can adapt.
- Add your own guardrails. Llama Guard, Prompt Guard, and Code Shield exist because the model is not meant to run in isolation.
- Evaluate on your own data. Scores from another model variant or a different year may not reflect the results you’ll achieve. If you are moving from model evaluation to building an autonomous AI application, our agentic AI roadmap covers agent design, tool use, guardrails, testing, and deployment.
- Mind the cutoff. Supply current facts in context, since the model’s knowledge ends in August 2024.
- Plan for lifecycle changes. Check each provider’s retirement dates before committing.
- Review attribution and naming obligations. The “Built with Llama” and “Llama” prefix rules apply to products and derived models.
Frequently Asked Questions
Q: What is LLaMA 4 Maverick?
A: It is Meta’s open-weight, natively multimodal mixture-of-experts model, released April 5, 2025, with 17B active and about 400B total parameters across 128 experts.
Q: How is Maverick different from Scout?
A: Both have 17B active parameters. Maverick has 128 experts and about 400B total parameters with a 1M context. Scout has 16 experts, about 109B total parameters, and a 10M context window.
Q: Is Llama 4 Maverick open source?
A: It is open-weight under the Llama 4 Community License. The license includes a 700 million user threshold, attribution rules, and an acceptable use policy, so it does not meet the Open Source Initiative’s definition.
Q: Can Maverick process images?
A: Yes. It accepts text and images and outputs text. The model card says image understanding was evaluated for up to five input images.
Q: What is its knowledge cutoff?
A: August 2024, according to the model card and cloud provider documentation.
Q: How big is the context window?
A: The model card lists 1 million tokens. Actual limits depend on the host, so check provider documentation.
Q: What hardware does it need?
A: Meta says the FP8 weights fit on a single H100 DGX host. Community estimates put weight memory around 800 GB at FP16, 400 GB at INT8, and 200 GB at INT4, before accounting for context cache.
Q: What was the LMArena controversy?
A: Meta reported a high LMArena score for an experimental chat-optimized variant that was not the released model. LMArena said Meta should have made that clearer and updated its policies. The released model ranked far lower when added.
Q: Can businesses in the European Union use it?
A: The Acceptable Use Policy does not grant the multimodal rights to individuals domiciled in, or companies based in, the European Union. Text-only uses are described differently by commentators, so seek legal advice.
Q: Is Maverick still available?
A: Yes, on several platforms. Amazon Bedrock marks it Legacy with an end-of-life date of March 30, 2027, and Meta has said its current Llama models will remain available as open source.
Conclusion and Final Thoughts
LLaMA 4 Maverick is a capable general-purpose, multimodal model whose reputation has been shaped as much by its launch as by its engineering. On Meta’s published numbers, it improved clearly over Llama 3.3 70B and reads charts and documents well for its era. Independent signals, though, place it far from the frontier on hard coding and expert reasoning, and the leaderboard dispute is a reminder to verify which model sits behind any score.
Three points are worth carrying away. Its memory demands follow the 400B total, not the 17B active count. Its license is workable for many businesses but is not open source, and it carries an EU multimodal carve-out that needs legal review. The provider’s documentation already outlines its lifecycle, so be prepared for future changes.
If you already run Maverick, the record gives no reason to panic, but it does suggest an evaluation against newer options. If you are starting fresh, compare current open-weight and hosted models on your own data before you pick one. Treat every number here, including the ones from Meta, as a starting point for your own measurement, and verify prices, limits, and dates on the provider pages that matter to you.