Llama-4-Maverick-17B-128E-Instruct

Llama 4 Maverick 17B-128E-Instruct is Meta's instruction-tuned, natively multimodal mixture-of-experts model from the Llama 4 family. It activates 17 billion parameters per forward pass across 128 routed experts (about 400 billion total parameters), uses early-fusion multimodal training, and supports text plus image inputs for assistant-style chat, visual reasoning, captioning, and code generation. Meta pretrained it on roughly 22 trillion multimodal tokens with an August 2024 knowledge cutoff and released it on April 5, 2025 under the Llama 4 Community License.

On widely reported benchmarks, Maverick ranks among the strongest open multimodal models in its class. Instruction-tuned scores include 80.5 on MMLU-Pro, 69.8 on GPQA-Diamond, 73.4 on MMMU, 43.4 on LiveCodeBench, and 77.6 on MBPP, while the pretrained checkpoint reaches 85.5 MMLU and 61.2 MATH. Meta reports that an experimental chat build reached an Arena Elo of 1417, and the model competes with GPT-4o and Gemini 2.0 Flash on multimodal tasks while matching much larger models on reasoning and coding.

The model ships with a 1,048,576-token context window, twelve supported languages plus broader pretraining coverage, and both BF16 and FP8 checkpoints on Hugging Face. FP8 weights are designed to fit on a single 8xH100 DGX node, making Maverick practical for high-quality multimodal deployment despite its large total parameter count.