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

This is the instruct variant of Llama 4 Maverick in our index: an open, multimodal-marked listing meant for people who want Meta's instruct weights rather than a raw or FP8-only repack. The name encodes a mixture-of-experts shape. The total parameter story and the active parameters are different ideas. Do not plan memory from the larger story alone.

Operators who want an open instruct model with image input and a long claimed window, and who will read Meta's license for acceptable use. You can run or rent GPUs and you will test the instruct chat format the card implies.

Skip the non-instruct and FP8 cards until this one is the model you meant. They are siblings. Skip it if Meta's license excludes your product. Our open-source badge will not argue with that license. Skip it for a pure text batch job if a smaller model you already serve is accurate enough. Multimodal capacity you do not use still costs memory.

1. Read the license and the prompt format upstream. Instruct models are picky about the template.
2. Run a text task and one image task. Grade them separately. A model can pass one and fail the other.
3. If memory is tight, then look at the FP8 row and repeat both tests. Adopt FP8 only if both grades hold.

Other open notes on this site are GLM-5.2, Kimi-K3, and DeepSeek-V4-Flash-0731, which are not drop-in templates. Hosted multimodal models such as Gemini Flash are the alternative when you will not serve weights. Multimodal is a filter, not a ranking.

Is 17B the full model? The name says 17B with 128 experts. Plan capacity from the runtime's reported memory, not from the 17 alone and not from a back-of-the-envelope total.

Which Maverick row should the config name? The one you tested. Copy the id from that card, including Instruct if that is what passed.