Meta · Meta AI (Llama)

MoE: 17B active params, 16 experts, 109B total. 10M-token context (longest of any released model); best for needle-in-haystack retrieval. Runs on a single H100.