by Meta · Llama 4 family · best for self-hosted multimodal workhorse with no vendor lock-in
Llama 4 Maverick is Meta's flagship open-weights model, released April 5, 2025 as the first Llama to ship as a Mixture-of-Experts. It pairs a 400-billion-parameter knowledge pool with only 17 billion active parameters per token, is natively multimodal (text + image from pre-training), and serves a 1M-token context window. The one-sentence buyer takeaway: it is not the smartest model on any leaderboard, but it is the strongest open-weights workhorse you can self-host and run on every major inference provider for roughly a tenth the price of a closed frontier API — making it the default hedge against vendor lock-in.
| Benchmark | Score | Source |
|---|---|---|
| MMLU | 85.5% | Meta / llm-stats aggregator2025-04-05T00:00:00.000Z |
| MMMU | 73.4% | Meta Llama 4 model card2025-04-05T00:00:00.000Z |
| MATH-500 | 61.2% | Meta (MATH-Hard)2025-04-05T00:00:00.000Z |
| MMLU-Pro | 80.5% | Meta Llama 4 model card2025-04-05T00:00:00.000Z |
| HumanEval | 85.8% | llm-stats aggregator2025-04-05T00:00:00.000Z |
| LMArena Elo | 1271 | LMArena (released Instruct; experimental chat ranked higher pre-release)2025-04-15T00:00:00.000Z |
| GPQA Diamond | 69.8% | Meta Llama 4 model card2025-04-05T00:00:00.000Z |
| LiveCodeBench | 43.4% | Meta Llama 4 model card2025-04-05T00:00:00.000Z |
| Aider Polyglot | 15.6% | Aider leaderboard (community)2025-04-10T00:00:00.000Z |
| Artificial Analysis Index | 18 | Artificial Analysis2026-05 |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“The smartest hedge in the market: open weights, every cloud, one-node deploy. I trade a little capability for total freedom from price hikes and lock-in.”
For a buyer weighing capability against vendor risk, Maverick is the strongest no-lock-in play in May 2026. Open weights plus availability on Bedrock, Vertex, Azure, and eight inference providers means you can never be held hostage on price or availability — the lesson every team that survived a closed-API price hike has internalized. Capability is good enough for the large majority of production workloads, and the sovereignty story satisfies EU and regulated-industry mandates. Roadmap confidence is the soft spot: Behemoth never shipped and Meta's Llama cadence is now uncertain, so treat Maverick as a durable present-day asset, not a guaranteed upgrade path.
“Maverick owns the 'open and multimodal on one node' square. Its moat is distribution and economics, not raw IQ — and that square is wide.”
Positioning-wise, Maverick wins where the market values control and cost over peak intelligence. Against closed frontier models it loses on capability; against other open weights (DeepSeek V3.x, Qwen 3) it competes on multimodality and the breadth of its provider ecosystem rather than benchmark wins. Its differentiation is being natively multimodal AND single-node deployable AND on every major cloud — a combination few open models match. Market timing is good: sovereignty and cost pressure are tailwinds. The risk is that DeepSeek and Qwen iterate faster, so Maverick's open-weights leadership is contestable.
“This is where the math flips. Ten-x cheaper to serve than closed frontier, and self-hosting pays back the GPUs in a quarter at volume.”
Maverick is the clearest TCO story in the lineup. On Bedrock it runs ~82–93% cheaper than Llama 3.1 405B for equal-or-better quality; DeepInfra floors at $0.15/$0.60 and Groq pushes per-call cost under a cent at chat lengths. The decision is API vs self-host: at sub-100M tokens/month, managed providers win on simplicity; above ~500M tokens/month on steady load, self-hosting on reserved 8xH100 typically amortizes the hardware in 3–5 months and then dominates. MoE keeps inference cheap relative to dense 400B. Watch the provider spread — Together's $2.19 output is 2–3x the floor, so naive provider choice can triple your bill.
“Portable weights, mature tooling, 1M context — but I still write evals because the chat template drifts between Groq, Bedrock, and Together.”
For hands-on builders, Maverick is genuinely portable: documented Hugging Face checkpoints, mature fine-tuning recipes (Together, Fireworks, Unsloth), and native support in vLLM, llama.cpp, Ollama, SGLang, and TensorRT-LLM. LoRA adapters land in hours. Structured output and JSON mode work but you often bolt on grammars/outlines yourself rather than relying on a first-class API. The recurring friction is provider drift — subtle chat-template and tool-format differences across hosts break agent loops if you do not test per provider. Tool use is reliable for single-step calls, shakier on long multi-step chains.
“Fast and fluent, never preachy — but the 'wow' moments belong to Claude and GPT-5. It's a solid backstage model, not a star.”
In daily use Maverick is competent and pleasant: fluent prose, sensible refusal rates, and sub-second first-token latency on Groq or Cerebras. Where it falls short is the top end — nuanced humor, emotional intelligence, and genuinely creative writing trail Claude and GPT-5 by a clear notch. For apps where the model sits behind a workflow, users never feel the gap. For a flagship chat product where the model IS the experience, the ceiling shows within an extended session. Long, context-heavy conversations also surface the comprehension degradation.
“Behemoth never shipped, the LMArena number came from a checkpoint you can't download, and 1M context is marketing once you test comprehension.”
Adversarially, Maverick has three claims to distrust. First, the splashy LMArena ranking used an unreleased "experimental chat" model; the actual Instruct weights rank materially lower — a textbook benchmark-presentation gap. Second, the 1M (and Scout's 10M) context is real for needle retrieval but Fiction.LiveBench ranks Llama 4 near the bottom on genuine long-context comprehension, so the headline number oversells usable capability. Third, the teacher model Behemoth (~2T params) was announced in April 2025 and still has not shipped amid reported capability concerns — the family's top end is vaporware. Real weaknesses: no reasoning mode, weak multi-file coding (Aider ~15.6), and a 2024 cutoff. The honest pitch is "good, cheap, open" — not "frontier."
Self-hosted or sovereign-cloud agent and RAG stacks where data must not leave the customer's infrastructure — Maverick is the strongest open option that runs on one node. Multilingual content pipelines across 100+ languages at high throughput. Document- and chart-heavy vision workflows where DocVQA/ChartQA-class accuracy matters. Cost-sensitive backends where a closed frontier API's per-token price or licensing is a non-starter and "good enough for 80% of production" is the bar.
The full research notes behind this review — verified against primary sources.
Maverick is a sparse MoE transformer: 400B total parameters across 128 routed experts plus one shared expert, with only ~17B parameters active per token (the shared expert plus one routed expert). MoE layers alternate with dense layers. Attention uses the iRoPE scheme — three of every four decoder layers apply rotary embeddings with chunked local attention (8K-token chunks), and every fourth "NoPE" layer drops positional encoding entirely and attends over the full context with temperature-scaled softmax to fight long-range probability fade. Vision is early-fused from a dedicated encoder into the same token stream. Meta discloses total/active params, expert count, training-token scale (30T+), and the 200-language mixture; exact layer counts, training compute, and the full data recipe are not fully published. The released checkpoint is Instruct-tuned; the FP8 variant fits an 8xH100 node.
Maverick scores as a strong-but-not-frontier generalist. Coding (cap_coding 6.5) is good for open weights — HumanEval ~85.8, LiveCodeBench 43.4 — but the Aider polyglot score (~15.6) exposes weakness on real multi-file edit workflows, so the editorial score sits mid-pack. Reasoning (6.0) and math (5.5) are competent on classic benchmarks (MMLU 85.5, MMLU-Pro 80.5, MATH-Hard ~61) but there is no native reasoning mode, so it loses badly to o-series, Claude extended thinking, and DeepSeek R1 on hard proofs and AIME-class problems. GPQA Diamond 69.8 is genuinely strong for an open non-reasoning model. Multilingual (8.0) is a real strength across 200 languages. Vision and document OCR (7.0 / 7.5) handle charts (ChartQA 85.3), documents (DocVQA 91.6), and MMMU 73.4 well. Long-context (4.5) is the honest weak spot: needle-in-haystack retrieval looks near-perfect, but comprehension benchmarks (Fiction.LiveBench) rank Llama 4 near the bottom, meaning the 1M window degrades on reasoning-over-context well before the ceiling. Function-calling and instruction-following (6.5 each) are reliable but trail the best closed models. No real-time data (0.0).
| Benchmark | Score | vs Predecessor (3.1 405B) | vs Top Competitor | Source |
|---|---|---|---|---|
| MMLU | 85.5 | -3.1 (405B 88.6) | ~ GPT-4o | Meta/llm-stats |
| MMLU-Pro | 80.5 | +7.2 (405B 73.3) | trails Claude/GPT-5 | Meta |
| GPQA Diamond | 69.8 | +18.7 (405B 51.1) | beats GPT-4o (53.6) | Meta |
| MATH (Hard) | 61.2 | comparable | trails reasoning models | Meta |
| HumanEval | 85.8 | ~ 405B (89.0) | ~ GPT-4o | llm-stats |
| LiveCodeBench | 43.4 | new | beats GPT-4o (~32) | Meta |
| Aider Polyglot | 15.6 | new | trails Qwen2.5-Coder-32B | Aider community |
| MMMU (vision) | 73.4 | new (405B no vision) | competitive | Meta |
| LMArena Elo | 1271 | n/a | mid-pack; experimental chat ranked higher | LMArena |
| Artificial Analysis Index | 18 | +1 (405B 17) | below reasoning frontier | AA |
Note the LMArena gap: Meta submitted an unreleased "experimental chat" checkpoint that ranked high; the actually-downloadable Instruct weights rank ~1271, materially lower. Treat any pre-release LMArena claim with suspicion.
Median output speed is ~104 tokens/sec across providers, with time-to-first-token around 0.66s on DeepInfra FP8. The MoE design means it generates roughly as fast as a dense 17B model despite the 400B pool. Specialty silicon pushes it far faster: SambaNova benchmarks ~645 tokens/sec, and Groq/Cerebras deliver sub-second interactive feel. For batch and high-throughput agent loops the economics are excellent; latency tier is fast.
| Surface | Cost | Notes |
|---|---|---|
| API input (representative) | ~$0.20 / 1M tok | DeepInfra/Fireworks FP8 floor |
| API output (representative) | ~$0.85 / 1M tok | spread $0.60–$2.19 across providers |
| DeepInfra | $0.15 in / $0.60 out | cheapest mainstream |
| Fireworks | $0.22 in / $0.88 out | |
| Groq | $0.59 in / $0.79 out | fastest interactive |
| Together | $0.55 in / $2.19 out | runs notably higher on output |
| SambaNova | blended ~$0.75 | 645 tps |
| AWS Bedrock | ~$0.50 in / ~$1.50 out | 82–93% cheaper than 3.1 405B on Bedrock |
| Self-hosted | 8xH100 node (FP8) | ~240GB VRAM to hold 400B at INT4 |
| Rate limits | provider-specific | no single Meta limit; managed tiers ~600 RPM / 200K TPM |
Open weights mean there is no single Meta price; the spread above is the inference-provider market as of May 2026. Together's output rate is an outlier 2–3x the floor — provider choice materially changes unit cost.
Open weights under the Llama 4 Community License. Download from Hugging Face (meta-llama/Llama-4-Maverick-17B-128E-Instruct, FP8 variant available). Self-hostable on a single 8xH100 node in FP8; INT4 quantization fits the full 400B parameter set in ~240GB VRAM (e.g. 4xH100 or an RTX 5090-class consumer box for aggressive quants). Managed availability spans AWS Bedrock, Google Vertex AI, Azure AI Foundry, OCI, and IBM watsonx. Inference providers include Together, Fireworks, Groq, DeepInfra, OpenRouter, Novita, SambaNova, and Hyperbolic — wide redundancy means you are never locked to one host. The Llama 4 Community License permits commercial use but requires a separate license from Meta if your products exceed 700 million monthly active users, and prohibits using outputs to train non-Llama models.
The weights ship without built-in content moderation; Meta provides Llama Guard 4 (a 12B natively-multimodal safety classifier for text + image, released April 2025) and Prompt Guard 2 (22M/86M jailbreak and prompt-injection detectors) as optional pre/post filters designed for the Llama 4 line. Because the model is self-hostable, "trains on inputs" is not applicable when you run it yourself; managed providers vary, but Meta's own terms do not train on your data. There are no model-level compliance certifications (SOC2/HIPAA/etc.) — those attach to whichever host or your own infrastructure. Refusal calibration is moderate and fully tunable via system prompt or fine-tune, which is precisely the point of open weights for regulated buyers. Governance falls under Meta's Frontier AI Framework.
First-class support across Hugging Face Transformers, vLLM, llama.cpp, Ollama, SGLang, TensorRT-LLM, and Unsloth, plus LangChain and LlamaIndex integrations. Available on Bedrock, Vertex AI, Azure AI Foundry, OCI, and IBM watsonx, and on Together, Fireworks, Groq, DeepInfra, OpenRouter, Novita, SambaNova, and Hyperbolic. Powers Meta's own consumer AI across WhatsApp, Instagram, and Messenger. Popularity is mainstream — the default open-weights multimodal pick for self-hosting teams in 2026.
There is no single Meta price; representative inference is ~$0.15–$0.59 input and ~$0.60–$2.19 output per 1M tokens depending on provider (DeepInfra cheapest, Together's output highest). Self-hosting trades per-token cost for GPU capex.
Yes. Download the FP8 weights from Hugging Face and serve on an 8xH100 node, or quantize to INT4 (~240GB VRAM) to fit smaller hardware.
Yes — image understanding is native (early fusion), not a bolted-on adapter, with strong DocVQA/ChartQA scores. It does not generate images.
For retrieval, largely yes; for reasoning across the full window, no — comprehension degrades well before the ceiling, so chunk and test on your workload.
The weights have no built-in moderation; add Llama Guard 4 / Prompt Guard 2. Compliance certifications come from your host or your own infra, not the model.
The Llama 4 Community License allows commercial use but requires a separate Meta license above 700M MAU and forbids training non-Llama models on its outputs.
Maverick for higher quality and 128 experts on a node; Scout for single-GPU deploy and the 10M context. Both share the same 17B-active speed profile.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-05-27