Meta Llama 4: Open-Weight Multimodal MoE Changes the Cost Equation

What shipped: Scout, Maverick, and where to find them

Meta released Llama 4 in two open-weight variants designed for different positions on the cost-capability curve. Llama 4 Scout is the efficient, volume-optimised model — lower memory footprint, faster inference, and well-suited to the long tail of requests where throughput matters more than peak capability. Llama 4 Maverick is the higher-capability sibling, carrying more total parameters and aimed at complex reasoning, multi-step analysis, and the harder prompts that Scout would handle less reliably.

Both models are available for free download on Hugging Face under the Llama community licence. Meta has also deployed them inside its own consumer products — WhatsApp, Messenger, Instagram Direct, and the Meta AI web interface — which means, unusually, the same open weights that power a mass-consumer product are available for builders to self-host. That consumer deployment at scale is itself a quality signal: Meta is betting its flagship AI assistant on these weights.

If you are coming from our earlier deployment how-to — which covers vLLM configuration, hardware sizing, and the practical serving setup — this article is the architecture and strategy layer above it. See the Llama 4 deployment guide for the infrastructure specifics.

Why MoE changes the inference maths

To understand why the Llama 4 architecture matters for builders, you need to understand the difference between a dense model and a Mixture-of-Experts model — because it is the difference between paying for a large model and paying for a much smaller one while getting quality closer to the large one.

In a dense transformer, every parameter fires on every token. If a model has 70 billion parameters, generating each output token involves compute proportional to those 70 billion parameters, multiplied by every layer. The total parameter count and the active compute per token are the same number.

A Mixture-of-Experts model breaks this coupling. The total parameters are spread across many specialised "expert" sublayers. For each token, a lightweight gating network inspects the token's representation and routes it to a small subset of experts — typically two to eight out of the full pool — while the rest remain dormant. The model accumulates the knowledge capacity of its full parameter count during training, but at inference time it only activates a fraction of that capacity per token.

The practical consequences for self-hosters are significant. A MoE model with a large total parameter count can deliver quality comparable to a dense model of similar capacity, while requiring substantially fewer FLOPs per generated token. Fewer FLOPs per token means lower GPU-hours per request, which means lower cost at a given level of quality. The trade-off is memory: you still need to hold the full parameter set in VRAM so the router can access any expert. This is why a MoE model with a large total count needs more memory than a dense model of the same active-compute budget, but less compute than a dense model of the same total count.

For Indian and UK teams deciding between self-hosting and paying an API, this architecture changes the break-even calculation. The cost per token to self-host a MoE model is lower than for a dense model of equivalent quality, which pushes the crossover point — the volume above which self-hosting becomes cheaper than the API — to a lower throughput threshold. At the GPU pricing available through India's IndiaAI Mission compute allocation or the UK's AI Research Resource, the maths can turn sharply in favour of self-hosting at volumes that would still favour an API with a dense model. For a detailed break-even analysis see our piece on AI inference cost economics in 2026.

Native multimodality: why baked-in beats bolted-on

Vision has been added to language models in two fundamentally different ways. The bolt-on approach takes a separately trained vision encoder — CLIP or a similar model — and attaches its output to the token stream that feeds a text-only language model. The native approach trains the model on paired image-text data from the start, so the representational space used for visual and textual signals is shared and jointly optimised.

Llama 4 Scout and Maverick use native multimodality. The visual understanding is not a module grafted on after the language model was already trained; it is part of the same pretraining run. This has concrete consequences for quality on tasks where text and image content are interdependent.

With a bolt-on vision model, the language backbone receives image features that have been encoded by a system that was never jointly optimised with it. The result is a loss of coherence on tasks where the meaning is in the relationship between text and image — not in either alone. Reading a receipt where the total is split across two lines, analysing a chart where the legend uses ambiguous colour coding, or parsing a form where the field labels are positioned above the answer boxes rather than beside them: these are the tasks where the seam between a separately trained encoder and a language backbone shows.

For builders in India and the UK, the practical implications cluster around document-intensive workflows. Several high-value use cases that are now commercially viable with native multimodality were unreliable or required extensive post-processing with bolt-on vision:

• Document parsing at scale — extracting structured data from scanned invoices, contracts, and government forms where layout is load-bearing. A native multimodal model reads the spatial relationships as well as the text.
• Form extraction for financial services — UK mortgage applications, Indian KYC documents, insurance claim forms. These mix printed and handwritten content with structured fields, and the meaning often depends on how elements are positioned on the page.
• Product image analysis for e-commerce — catalogue enrichment, attribute extraction, and defect detection where the textual description and the image need to be understood together, not in parallel pipelines that are stitched at the end.
• Medical and diagnostic imaging with clinical text — a domain where the correlation between what the image shows and what the accompanying report says is exactly the question being asked, and where bolt-on vision's seam is most costly.

The important caveat is that native multimodality is a necessary but not sufficient condition for quality on these tasks. Model capability, fine-tuning, and prompt engineering still matter. But it removes a structural impediment that limited what bolt-on vision could achieve regardless of how much effort you put into the other variables.

The refusal rate story: from 7% to under 2%

Llama 3.3 had an over-refusal problem. On social and political topics, Meta's own measurements found a refusal rate of around 7% — meaning the model declined to engage with roughly one in fourteen requests in those categories. That sounds like a safety feature, but over-refusal has real costs for product builders: failed requests that should have succeeded, frustrated users who get a refusal when they asked a legitimate question, and engineering time spent on prompt workarounds that should not be necessary.

Llama 4 Scout and Maverick bring that figure below 2% per Meta's published figures. For context, the goal is not zero refusals — a model that refuses nothing would be genuinely dangerous. The goal is calibrated refusals: declining clearly harmful requests while engaging with legitimate ones. Moving from 7% to under 2% is a significant shift toward calibration.

For builders, the practical consequences are several. Fewer edge-case failures in customer-facing conversational flows means fewer fallback paths and less defensive prompt engineering. The long tail of legitimate requests that previously bounced off a refusal — questions about contested historical events, sensitive policy topics, comparative analysis of ideologies — now resolve. This matters most for builders in news summarisation, educational technology, legal research assistance, and public-sector information services, where the topics are often sensitive by definition and a 7% refusal rate is a product-level problem.

The reduction also changes the fine-tuning calculus. When you fine-tune a base model that over-refuses, you inherit that tendency and often have to work against it. A model that starts with a more calibrated refusal behaviour is a better base for domain adaptation.

Behemoth as the teacher: knowledge distillation at scale

Scout and Maverick were not trained in isolation. Meta has a third model in the Llama 4 family — Llama 4 Behemoth — which functions as the teacher in a knowledge distillation pipeline. Behemoth is not publicly released; it is Meta's internal frontier model that is used to generate high-quality training signal for the smaller, deployable models.

Per Meta's announcement, Behemoth outperforms GPT-4.5 and Claude Sonnet 3.7 on MATH-500 and GPQA Diamond benchmarks. Those are hard mathematical reasoning and graduate-level science question benchmarks respectively — evaluations where the signal is not easily gamed and where the capability gap between models is meaningful. The precise scores are Meta's published figures; treat them as directional rather than as definitive cross-lab comparisons, since benchmark conditions and evaluation details vary.

What matters for Scout and Maverick is what leaks through the distillation. Knowledge distillation is not lossless — a student model trained on a teacher's outputs cannot fully replicate the teacher's capability. But the training signal quality from a frontier-level teacher has a measurable effect on the student's reasoning behaviour, particularly on the multi-step problems where the teacher's chain of thought provides the richest learning signal. The mathematical reasoning calibration in Scout and Maverick reflects, in part, the quality of reasoning Behemoth can demonstrate.

For builders, the distillation story matters because it explains why Llama 4 models punch above what their size alone would predict. The capability is not just a function of total parameters — it is a function of the training data quality and the teacher signal behind it.

Head-to-head context: Llama 4 in the June 2026 open-weight landscape

Llama 4 lands into a field that has moved considerably since the original Llama 4 announcement period. The June 2026 open-weight landscape includes several strong competitors that builders should have in their evaluation set alongside Scout and Maverick.

DeepSeek V4 Pro holds the aggregate open-weight leaderboard position on the Artificial Analysis Intelligence Index — see our DeepSeek V4 Pro leaderboard analysis for the detail. It ships under MIT, with a 1M-token context window and a MoE architecture of its own. Where it differs from Llama 4 is in the multimodal dimension: DeepSeek V4 Pro is primarily text-focused at the frontier, while Llama 4 Scout and Maverick have native image understanding baked in from pretraining.

Gemma 4 from Google is the other major Apache-licensed open-weight multimodal contender — a 27B total-parameter MoE that activates only 4B parameters per token, with configurable chain-of-thought depth and built-in function calling. See our Gemma 4 deep dive for the comparison. Gemma 4's advantage is the permissive Apache-2.0 licence, which gives legal teams an easier path than the Llama community licence. Llama 4's advantage is the Behemoth-distilled reasoning quality and the native multimodal coherence.

GLM-4.7 from Zhipu AI is notable for a reported near-zero hallucination rate on structured factual queries — a claim that is worth testing on your own workload. See our GLM-4.7 analysis. And for the broader June 2026 coding benchmark picture — where MiniMax M3, Kimi K2 variants, and the Chinese open-weight cohort are all competitive — see the June 2026 coding shoot-out.

The headline for builders in mid-2026: the open-weight field is genuinely competitive. No single model dominates across all dimensions. The right model depends on your capability priorities, your licence requirements, your multimodal needs, and your serving constraints. Run your own evals on your task mix — the leaderboard is a shortlist tool, not a verdict.

Scout vs Maverick: the comparison table

Dimension	Llama 4 Scout	Llama 4 Maverick
Architecture	MoE, smaller total parameters, fewer active per token	MoE, larger total parameters, more active per token
Multimodality	Native (text + image from pretraining)	Native (text + image from pretraining)
Refusal rate	Under 2% (per Meta)	Under 2% (per Meta)
Licence	Llama community licence	Llama community licence
Availability	Hugging Face (free download)	Hugging Face (free download)
Inference cost tier	Lower — suited to high-throughput, cost-sensitive routes	Higher — suited to quality-critical, lower-volume tasks
Primary use cases	Document parsing, form extraction, high-volume Q&A, summarisation	Complex reasoning, multi-image analysis, agentic tasks, hard coding problems
Best fit	Bootstrapped startups; price-sensitive volume workloads	Enterprises; quality-demanding internal tools; research

Builder decision matrix: Scout vs Maverick vs proprietary API

The decision between Scout, Maverick, and a proprietary frontier API is not a single question — it is three separate questions answered in sequence. First, does self-hosting make sense at all? Second, if so, which open-weight model? Third, are there workloads that should stay on a closed API regardless?

When self-hosting is the right call: Data residency requirements that prevent personal or regulated data from leaving your infrastructure. High steady-state volume where the per-token API cost exceeds the amortised compute cost. Deep fine-tuning needs that require access to the weights. Reproducibility requirements where you cannot accept the model changing underneath you on a third party's release schedule.

When to choose Scout over Maverick: Your primary workload is document understanding, structured extraction, or high-volume Q&A where Scout's quality is sufficient. Your GPU budget is constrained and you need to maximise requests per hour. You are building a latency-sensitive product where Scout's faster inference matters. You plan to route the long tail of easy requests through Scout and only escalate hard tickets to Maverick or a proprietary model.

When to choose Maverick over Scout: Your workload involves complex multi-step reasoning, longer-horizon agentic tasks, or multi-image analysis where the capability difference is material. Quality on hard prompts is more important than cost per token. You have the hardware headroom to absorb Maverick's larger serving footprint.

When to stay on a proprietary API: You need the absolute top of the capability curve for a small volume of high-stakes requests, and the incremental quality is worth the cost. Your traffic is spiky and you do not want to maintain idle GPU capacity for burst headroom. You are early in your product lifecycle and want to defer the operational overhead of running a serving stack until you have validated product-market fit.

The most resilient architecture in 2026 is the hybrid: Scout or Maverick for the high-volume, steady-state workloads where the economics favour self-hosting; a closed frontier API for the low-volume, capability-maximal, or burst cases. See our 2026 inference cost economics piece for the full break-even methodology.

India and UK builder angles: why zero licensing cost changes the maths

The Llama 4 release has different implications depending on where you are building. Both India and the UK have structural reasons to care about open-weight multimodal models that go beyond the generic "free weights" headline.

For bootstrapped Indian startups, the zero licensing cost of Llama 4 is significant at the stage where every rupee of operating cost is scrutinised. A proprietary multimodal API — particularly one priced for Western enterprise customers — can consume a disproportionate share of an early-stage product's margin. Running Scout on the IndiaAI Mission's subsidised compute converts that variable per-call cost into a much lower fixed infrastructure cost. For a product doing document understanding at Indian-SME scale — digitising paper-based records, processing vernacular-language forms, extracting data from handwritten ledgers — the combination of native multimodality and near-zero per-token cost is a genuine competitive enabler rather than a marginal optimisation.

The vernacular language dimension matters here too. Per Meta, the model supports multiple Indian languages, which means the multimodal pipeline handles mixed-script documents — a page that is partly in English and partly in Hindi, Tamil, or Bengali — without the quality degradation that many Western-trained models show on non-Latin scripts. For builders serving non-English-speaking Indian users, this is a functional requirement, not a nice-to-have.

For UK enterprises, the open-weight story is primarily about auditability and compliance. Under UK GDPR, data controllers must be able to account for automated decision-making that affects individuals. A closed API is a black box — you have contractual data-handling commitments, but you cannot inspect what the model does with the data or verify the model's behaviour directly. Open weights allow your team to audit the model's inference behaviour, to run it inside your own controlled infrastructure, and to maintain the technical documentation that regulators increasingly expect for AI systems used in financial services, healthcare, and public-sector contexts.

The UK's AI Security Institute (formerly the AI Safety Institute) has also signalled that organisations deploying AI in high-risk categories should be able to demonstrate understanding of and control over the systems they use. Open weights give UK enterprises a credible path to that demonstration in a way that a closed API does not, regardless of the API provider's enterprise compliance programme.

For fine-tuning guidance specific to open-weight models on constrained budgets — which applies to both bootstrapped Indian startups and UK teams with limited ML infrastructure — see our 2026 guide to fine-tuning with LoRA and QLoRA.

What to watch next

Llama 4 Behemoth remains unreleased to the public. Meta has not indicated a timeline, and it may remain internal indefinitely as a training resource rather than a deployable model. If it does ship as open weights, it would immediately be among the most capable open-weight models available — but that is speculative, and builders should plan against what is available today.

The more immediate question for the open-weight landscape is whether the multimodal architecture in Scout and Maverick — native vision, MoE, under 2% refusal rate — becomes the expected baseline for open-weight releases in the second half of 2026. Given the pace at which the field has moved — Gemma 4, GLM-4.7, MiniMax M3, DeepSeek V4 Pro all shipping within a few weeks of one another — the competitive response to Llama 4's multimodal positioning is likely already in progress at multiple labs.

For builders, the right response to this pace is not to wait for the dust to settle. The dust does not settle in this field. Build your evaluation harness now, run your task-representative golden set against each competitive release, and route workloads based on your own measurements rather than aggregate benchmark rankings. The April 2026 open-weight landscape review and the June 2026 coding shoot-out give you the competitive context; your evals give you the routing decision.

Source: Meta AI blog announcement. Benchmark figures (MATH-500, GPQA Diamond) are per Meta's published claims; always verify benchmark conditions and current leaderboard standings independently before using them to make production decisions.