A 400B-parameter natively multimodal Mixture-of-Experts model with 17B active parameters, 128 routed experts and a documented 1M-token base context window.
This passport separates the base model from its legal conditions and from each deployment implementation. That distinction matters for Maverick: weight access is broad but not unrestricted, the multimodal Llama 4 policy contains an EU developer restriction, and real context or feature limits differ substantially across self-hosted and managed deployments.
Llama 4 Maverick is a natively multimodal open-weight AI model developed by Meta as part of the Llama 4 model family. It is designed to process both text and images and to generate multilingual text and code.
Maverick uses a Mixture-of-Experts (MoE) architecture with approximately 400 billion total parameters, while about 17 billion parameters are active for each token. The model contains 128 routed experts plus a shared expert and supports a documented context window of up to 1 million tokens.
Unlike a conventional dense 17B model, Llama 4 Maverick still requires infrastructure capable of storing and serving its much larger total parameter set. The sparse MoE architecture reduces the amount of computation used for each token, but it does not reduce the complete model to the hardware footprint of a 17B model.
Meta released Llama 4 Maverick on 5 April 2025 as one of the first Llama models combining native multimodality with a Mixture-of-Experts architecture. Official weights are available for independent deployment under the Llama 4 Community License Agreement.
In practical terms, Llama 4 Maverick is intended for workloads such as multimodal assistants, image understanding, document analysis, coding, multilingual generation, long-context processing and AI systems that require self-hosted or managed deployment options.
Meta Llama 4 model card ↗Llama 4 Maverick combines the storage footprint of a very large model with the per-token compute characteristics of a sparse Mixture-of-Experts system. It is therefore misleading to treat “17B active parameters” as if Maverick were a conventional 17B model.
Meta documents approximately 400 billion total parameters, while only about 17 billion parameters are activated for an individual token. The MoE layers contain 128 routed experts plus a shared expert; each token passes through the shared expert and one routed expert. The sparse route lowers active computation, but the whole parameter set still has to be available to the serving system. That is why deployment remains datacenter-class even though the active count looks comparatively small.
Maverick is also natively multimodal. Meta uses early fusion so image and text tokens are processed in a unified model backbone rather than attaching vision only as a late external component. The official model card lists multilingual text and image input, multilingual text and code output, a 1M context window and an August 2024 knowledge cutoff.
Legally, Maverick is not an Apache-2.0- or MIT-style release. It is distributed under the Llama 4 Community License Agreement, with an incorporated Acceptable Use Policy. The license grants broad rights to use, modify and redistribute subject to conditions, while the policy adds prohibited-use rules and, for Llama 4 multimodal models, an important EU restriction for certain developers. This passport therefore marks commercial use as conditional, not simply “yes”.
Meta model card ↗Meta architecture announcement ↗These values describe the base Maverick model unless a deployment provider is explicitly named. Provider-specific limits appear later and should not be silently inherited from the base model.
Meta describes Llama 4 as an autoregressive architecture using alternating dense and Mixture-of-Experts layers. In Maverick's MoE layers, routing sends each token through a shared expert and one of 128 routed experts. The design aims to preserve the representational capacity of a very large parameter pool without activating that entire pool for every token.
This distinction affects both performance and infrastructure. Active parameters are relevant to compute performed per token, but total parameters remain relevant to weight storage, model loading, sharding and memory topology. A capacity plan based solely on 17B would therefore substantially understate Maverick's deployment footprint.
Only a subset of parameters participates in each token's forward pass. This reduces active compute relative to a dense model with the same total parameter count, but it does not eliminate the need to store and serve the complete expert set.
Text and visual tokens are integrated into the same model backbone. Meta says its vision encoder is based on MetaCLIP and was adapted in conjunction with a frozen Llama model before multimodal joint training.
Meta documents 1M tokens for Maverick. Long context has direct implications for KV-cache memory, latency and serving economics, so the model maximum should never be treated as a cost-free default operating point.
Meta reports roughly 22 trillion multimodal pretraining tokens for Maverick from a mixture of publicly available data, licensed data and information from Meta products and services, including publicly shared Instagram and Facebook posts and interactions with Meta AI. The model card lists August 2024 as the data freshness cutoff.
The broader Llama 4 pretraining effort covered 200 languages, with more than 100 languages receiving over one billion tokens each. Meta explicitly lists 12 languages as supported for Maverick deployments, so broader pretraining coverage should not be interpreted as equivalent support quality.
For post-training, Meta describes a pipeline of lightweight supervised fine-tuning, online reinforcement learning and lightweight direct preference optimization. It also says Maverick was co-distilled from the larger Llama 4 Behemoth teacher during pretraining.
~22T multimodal tokens; text, image and video-related data; broad multilingual corpus. Meta says Llama 4 models were pretrained with up to 48 images in multimodal sequences.
Lightweight SFT → online RL → lightweight DPO, with a curriculum intended to balance multimodal, reasoning and conversational capability.
Meta reports co-distillation from Llama 4 Behemoth, a much larger teacher model, to improve end-task quality while keeping Maverick's active computation lower.
Capabilities below separate base-model intent from provider-added features. Tool calling, structured outputs and guardrails can change depending on the serving platform even though the underlying weights remain the same.
The Instruct checkpoint is intended for assistant-like chat, generation, summarization, multilingual workflows and general instruction following.
Meta lists visual recognition, image reasoning, captioning and visual question answering. The model is text-output only; it does not generate images.
Meta's model card says the family was tested for image understanding with up to five input images and recommends additional application-specific testing beyond documented conditions.
The model card includes code benchmarks and describes multilingual text and code output. Coding capability is part of the core model rather than a cloud-only extension.
The 1M base context supports large-document and code-context scenarios, but provider implementations may expose lower practical limits and long contexts increase memory and latency costs.
Do not treat tool calling as a universal Maverick property. AWS documents client-side tool calling and Agents; Google documents function calling; current Microsoft Foundry documentation lists tool calling as unavailable for its Maverick offering.
The table below is intentionally limited to a small set of measurements from Meta's official model card. Meta states that its reported evaluations were conducted on BF16 models even though quantized checkpoints are also available. Scores should not be treated as universal real-world performance guarantees.
| Area | Benchmark | Metric | Maverick | Interpretation |
|---|---|---|---|---|
| Reasoning & knowledge | MMLU Pro | macro_avg / accuracy | 80.5 | Instruction-tuned, 0-shot in Meta's evaluation. |
| Science reasoning | GPQA Diamond | accuracy | 69.8 | High-difficulty factual/reasoning benchmark. |
| Coding | LiveCodeBench | pass@1 | 43.4 | Meta reports a dated evaluation window; benchmark versions matter. |
| Image reasoning | MMMU | accuracy | 73.4 | Measures multimodal reasoning across academic domains. |
| Image understanding | ChartQA | relaxed accuracy | 90.0 | Chart-oriented visual question answering. |
| Multilingual math | MGSM | average / exact match | 92.3 | Multilingual grade-school math evaluation. |
Maverick uses the Llama 4 Community License Agreement, not a standard permissive software license. Meta grants a limited, worldwide, royalty-free license to use, reproduce, distribute, create derivative works and modify the Llama Materials, but redistribution, attribution, scale and policy conditions apply.
The Acceptable Use Policy is explicitly incorporated into the agreement. OpenWeightModels therefore treats the license document and the policy document as two separate but jointly relevant sources.
Custom commercial/community license · effective 5 April 2025 · accompanied by incorporated Acceptable Use Policy.
| Question | Classification | What the source says in practice |
|---|---|---|
| Access / use weights | Allowed* | License grants rights to use Llama Materials, subject to the agreement and incorporated policy. |
| Commercial use | Conditional | Commercial/research use is intended, but scale, policy and regional conditions can materially affect eligibility. |
| Modification / fine-tuning | Allowed* | Modification and derivative works are included in the grant, subject to the terms. |
| Redistribution | Conditional | Agreement copy, “Built with Llama” display and Notice-file attribution requirements apply when distributing/making materials or containing products/services available. |
| Derived AI model naming | Conditional | If Llama materials or outputs are used to create/train/fine-tune/improve another AI model that is distributed or made available, the agreement requires “Llama” at the beginning of that model name. |
| 700M MAU clause | Special term | If the licensee or affiliates exceeded 700M monthly active users in the month preceding the Llama 4 release, a separate Meta license is required before exercising rights. |
| Acceptable Use Policy | Mandatory | The policy is incorporated by reference and contains prohibited-use categories plus the multimodal EU developer restriction. |
| OSI-approved open source | No standard OSS license | OpenWeightModels classifies the weights as open-weight but does not classify the license as a standard permissive OSI software license. |
The phrase “supports 1M context” is incomplete without naming the implementation. Maverick's base model is documented at 1M, NVIDIA NIM supports 1M on H200 but 430K on H100, Google currently documents 524,288 on its managed endpoint, while AWS documents 1M. Provider-specific features such as tool calling also differ.
OpenWeightModels therefore stores deployment facts as child records of the model rather than flattening them into a single global feature list.
| Route | Control | Context / output | Modalities | Notable feature |
|---|---|---|---|---|
| Direct BF16 / FP8 weights | Highest | Base model: 1M documented | Text + image → text | Full operator responsibility for serving, memory and policy compliance. |
| Hugging Face + vLLM/SGLang | High | Implementation-dependent | Text + image → text | Official model page documents Transformers, vLLM and SGLang launch paths. |
| NVIDIA NIM | High / packaged | H200: 1M · H100: 430K | Vision-language serving | Official FP8 checkpoint only; validated 8-GPU node configurations. |
| Amazon Bedrock | Managed | 1M context · 8K max output | Text + image → text | Streaming, Guardrails, client-side tool calling, Flows and Agents documented. |
| Google managed endpoint | Managed | 524,288 context · 8,192 output | Text/code/image → text | Function calling and structured output documented; model availability in us-east5. |
| Microsoft Foundry | Managed | 1M documented input context | Text + images → text | Global Standard deployment; current Microsoft listing says tool calling: No. |
Meta releases both BF16 and official FP8 Maverick checkpoints. Direct deployment offers the most control but shifts model serving, scaling, observability, safeguards and legal compliance to the operator.
The official model page documents Transformers loading plus vLLM and SGLang serving, including OpenAI-compatible API examples for the serving frameworks.
NVIDIA validates Maverick only on homogeneous eight-GPU H100/H200 nodes in its current NIM support matrix. Generic configurations are not supported for this profile.
AWS lists Maverick as an active Bedrock model with text/image input and text output. AWS also applies geofencing to Llama 4 access based on account country and request source.
Google's current managed Maverick record exposes a materially smaller context envelope than Meta's base-model maximum and adds platform features such as function calling and structured output.
Microsoft currently lists the FP8 Maverick Instruct model as a Foundry model sold directly by Azure with Global Standard deployment. Its current capability table does not advertise tool calling for this model.
Meta releases Maverick in BF16 and FP8. It states that the FP8 weights can fit on a single H100 DGX host while maintaining quality. NVIDIA's production NIM profile is more explicit: it supports a node of eight H100 SXM, H100 NVL, H200 SXM or H200 NVL GPUs and uses Meta's official FP8 checkpoint only.
Long context further changes memory requirements because KV-cache usage grows with sequence length, batch size and implementation. For this reason, OpenWeightModels does not publish one universal “minimum VRAM” number for Maverick.
Official Meta weight format and the precision Meta says it used for its reported evaluations. It provides a high-quality reference point but carries the largest raw weight-memory requirement of the official releases.
Official quantized release. Meta says the FP8 weights fit on one H100 DGX host. NVIDIA NIM uses this official FP8 checkpoint for its validated Maverick serving profile.
Meta also provides code for on-the-fly INT4 quantization. That is distinct from saying every INT4 community artifact has identical behavior, provenance or quality.
Model-family names are not deployment identifiers. Precision and tuning state matter for reproducibility, memory planning and comparison.
| Variant | Purpose | Precision | Status | Notes |
|---|---|---|---|---|
| Llama-4-Maverick-17B-128E-Original | Original/pretrained-style release | BF16 | Official | Useful when adapting the base model rather than consuming the instruction-tuned behavior. |
| Llama-4-Maverick-17B-128E-Instruct | Assistant / visual reasoning | BF16 | Official | Primary instruction-tuned checkpoint documented for chat and visual reasoning. |
| Llama-4-Maverick-17B-128E-Instruct-FP8 | Reduced-memory serving | FP8 | Official | Official quantized release; used by NVIDIA NIM's validated Maverick profile. |
| Community GGUF / other quantizations | Alternative runtimes | Varies | Community | Should be tracked separately by publisher, quantization recipe and source; not equivalent to Meta's official checkpoint. |
A useful model passport should make deployment risks and interpretive limits as visible as capabilities.
The 17B figure is the active parameter count. Maverick still contains roughly 400B total parameters, making model storage and orchestration a large-system problem.
The base model is documented at 1M, but Google exposes 524,288 on its managed endpoint and NVIDIA limits H100 NIM deployments to 430K. Provider context must be recorded separately.
AWS and Google expose tool/function-oriented capabilities around their Maverick services; current Microsoft Foundry documentation lists tool calling as unavailable for its Maverick model.
The incorporated AUP contains a specific EU developer restriction for Llama 4 multimodal models while carving out end users of incorporated products/services. This deserves explicit legal review for EU organizations.
Meta lists August 2024 as the knowledge cutoff. Retrieval or other external-data architecture may be needed for newer factual information.
Meta says its reported evaluations were run on BF16 models. Provider quantization, serving parameters, prompts, safety layers and context lengths can change application-level behavior.
Last full review: 29 September 2026. Architecture, training, benchmark and weight claims are grounded primarily in Meta sources. Deployment-specific claims are grounded in the documentation of the named platform rather than inferred from the base model.
Material source changes should update both the affected field and this history instead of silently overwriting the record.
Gold-standard passport created. Expanded architecture, training, capabilities, benchmarks, detailed license analysis, provider-level deployment records, hardware/quantization, variants and limitations. EU restriction source corrected to the incorporated Acceptable Use Policy rather than the core license text.
Initial Model Passport created with license and deployment overview.
OpenWeightModels calls it open-weight because official weights are available. Its custom Llama 4 Community License and incorporated use policy differ from standard permissive software licenses, so “open weight” is the more precise classification for this registry.
Meta's model card intends Llama 4 for commercial and research use, but the answer is conditional on the Community License and AUP. Important conditions include redistribution obligations, the 700M-MAU clause and the multimodal EU developer restriction.
Do not confuse a single H100 GPU with an H100 DGX host. Meta says the FP8 checkpoint fits on a single H100 DGX host; NVIDIA's validated NIM profile uses eight H100 or H200 GPUs.
Meta documents a base context window of up to 1 million tokens for Llama 4 Maverick. Provider implementations can expose smaller limits: Google's current managed Maverick endpoint documents 524,288 tokens, while NVIDIA NIM documents 430,000 on H100 and 1 million on H200.
There is no single provider-independent yes/no answer in practice. AWS documents client-side tool calling, Google documents function calling, while current Microsoft Foundry documentation lists tool calling as unavailable for its Maverick offering.