A reasoning-specialist dense model, a general-purpose open-weights workhorse, and a managed-API flagship. DeepSeek R1 Distill Llama 70B takes a Llama 3.3 70B base and applies chain-of-thought supervision distilled from the full 671B R1 MoE, released January 2025. Independent evaluations place it at roughly 70–80 percent of full R1 performance on AIME and MATH benchmarks at significantly lower inference cost. MIT license means no commercial friction. Groq, DeepInfra, and Fireworks all carry it, with Groq offering particularly low latency at this parameter count. If your bottleneck is multi-step mathematical or logical reasoning and you want a 70B-class cost profile, this is the obvious pick. Llama 3.3 70B Instruct is Meta's December 2024 general-purpose model — same parameter count, 131K context, Llama 3 community license. It outperforms the R1 distill variant on open-ended instruction following and creative tasks where explicit chain-of-thought is not the right approach. The breadth of provider coverage and permissive licensing make it the lowest-friction starting point for new projects that do not have strong reasoning requirements. Mistral Large 2 runs at 123 billion parameters, which gives it a quality ceiling noticeably above either 70B option. It shines on multilingual structured output, function calling, and European-language quality, but costs more and requires Mistral's managed API for most production deployments under the Research License. Pick DeepSeek R1 Distill 70B for math-heavy agents, code reasoning, or any pipeline where chain-of-thought traces are an asset. Pick Llama 3.3 70B for general workloads where Apache licensing and provider flexibility come first. Pick Mistral Large 2 when you need higher parameter capacity and are comfortable with Mistral's API ecosystem.