Llama 4 Maverick

Llama 4 Maverick

Quick answer: Llama 4 Maverick is Meta's April 2025 frontier open-weight multimodal MoE model with 17B active parameters (from a 400B+ total MoE). It scores 93.8% on the AI2 Reasoning Challenge, 80.5% on MMLU-Pro, and 5.68% on Humanity's Last Exam, with a 1M-token context window. Available under the Llama 4 License with open weights on Hugging Face.

At a Glance

Where Llama 4 Maverick leads

• 93.8% AI2 Reasoning Challenge — strong multi-discipline scientific reasoning
• 80.5% MMLU-Pro — competitive with GPT-4 class proprietary models
• 1M-token context window — matches GPT-4.1 and Gemini 2.5 Pro
• Open weights — self-hostable for cost-sensitive high-volume deployments
• Multimodal: text and image inputs
• MoE efficiency: 17B active params at inference from 400B+ total

Where it lags

• 5.68% HLE — below frontier models (GPT-5.6 Sol, Gemini 2.5 Pro)
• 28.4% BigCodeBench — moderate coding performance
• Llama 4 License (not Apache 2.0) — some commercial use restrictions
• Requires significant infrastructure to self-host the full MoE

Best for: Open-weight deployments needing multimodal + 1M context at low inference cost; research use; organisations requiring data sovereignty.

What Llama 4 Maverick Is

Llama 4 Maverick is Meta's highest-capability model in the Llama 4 family, released April 5, 2025 alongside Llama 4 Scout. Both are multimodal Mixture of Experts (MoE) models using 17B active parameters from a much larger parameter pool, making inference efficient relative to their knowledge capacity.

Maverick's headline feature is its combination of 1M context window, multimodal capability, and open weights. This makes it the most capable open-weight option for long-context tasks at launch, competing directly with GPT-4.1 and Gemini 2.5 Pro on context length while remaining freely available for self-hosting.

At 80.5% MMLU-Pro, Maverick matches or exceeds many proprietary models launched in 2024. Its 93.8% AI2 RC score indicates strong factual scientific reasoning. For organisations that require data sovereignty or high-volume cost efficiency with open weights, Llama 4 Maverick is the strongest available option in the Llama 4 family.

Specifications

Pricing

Llama 4 Maverick is available as open weights — free to download and self-host. Hosted API access is available via Meta AI and third-party providers (Together AI, Groq, etc.) at market rates.

Context Window

Llama 4 Maverick has a 1,000,000-token context window — matching GPT-4.1 and Gemini 2.5 Pro in the largest-context tier.

Public Benchmark Scores

Llama 4 Maverick vs Alternatives

Llama 4 Maverick vs Scout: Maverick has higher MMLU-Pro (80.5% vs 74.3%) and AI2 RC (93.8% vs 90.7%); Scout has a 10× larger context window (10M vs 1M). For most tasks requiring open weights and large context, Maverick is the better performer.

Run Llama 4 Maverick

# Via Together AI (hosted)
from together import Together

client = Together(api_key="YOUR_API_KEY")
response = client.chat.completions.create(
    model="meta-llama/Llama-4-Maverick-17B-128E-Instruct",
    messages=[{"role": "user", "content": "Analyse this dataset..."}],
)
print(response.choices[0].message.content)

Related Models & Benchmarks

• Compare: Llama 4 Scout · GPT-4.1 · Gemini 2.5 Flash
• Benchmarks: AI2 Reasoning Challenge · MMLU-Pro · Humanity's Last Exam

Frequently Asked Questions

Specs from Meta's official Llama 4 announcement (April 5, 2025) and Benchgen evaluations. Last updated 2026-07-24.