253B (253,401,268,224 parameters; dense)
128K (131,072 tokens)
NVIDIA Open Model License + underlying Llama 3.1 Community License Agreement and Acceptable Use Policy; commercial use subject to terms (not Apache 2.0)
Text-only reasoning and chat model derived from Meta Llama-3.1-405B-Instruct, not Llama 4. Puzzle neural architecture search compresses the dense decoder with skipped attention, variable-width FFNs and FFN fusion; recovery uses 65B knowledge-distillation tokens followed by 88B continual-pretraining tokens. Supervised fine-tuning covers math, code, reasoning, chat and tool calling, followed by multiple GRPO reinforcement-learning stages and alignment. Reasoning is toggled with the system prompt detailed thinking on/off; NVIDIA recommends placing other instructions in the user prompt. Intended for agents, RAG, tool use and English/code, with additional supported languages. NVIDIA reports reasoning-on scores of 76.01 GPQA, 72.50 AIME25, 66.31 LiveCodeBench (2024-08-01 to 2025-02-01), 97.00 MATH500 and 74.10 BFCL V2 Live, evaluated at 32K sequence length; reasoning-off scores are substantially lower on math, science and coding. The original downloadable checkpoint is BF16 and NVIDIA lists 8x H100-80GB or 4x B100 test hardware; memory, KV-cache and serving settings limit practical concurrency/context. The separate CPT base checkpoint and later official FP8 packaging are not grouped into this launch card.