DeepSeek V3 and R1 Models: Open-Weights Self-Hosting vs API for Enterprise Math and Code in 2026

Overview of DeepSeek's V3 and R1 Model Family DeepSeek has established itself as a leader in open and accessible large language models (LLMs), particularly with its V3 and R1 families, optimized for enterprise operations involving math, code, reasoning, and agentic workflows. As of May 4, 2026 (UTC), DeepSeek's official documentation at platform.deepseek.com and api-docs.deepseek.com highlights the V3 series for efficient general-purpose tasks and the R1 family for advanced reasoning, including successors like V3.2, V4-Flash, V4-Pro, and R1-Distill variants. The V3 family, such as and , emphasizes speed and cost-efficiency with mixture-of-experts (MoE) architectures for standard completions and chat. In contrast, the R1 family—featuring , , and distilled models like and —boasts a 671B parameter MoE setup with only 37B activated parameters per token, supporting 128K context lengths for co

mplex step-by-step reasoning. These models outperform predecessors on math and code benchmarks, with R1-Distill variants achieving state-of-the-art results comparable to closed models like OpenAI's o1-mini, per DeepSeek's GitHub repository (github.com/deepseek-ai/DeepSeek-R1). DeepSeek notes deprecation of legacy models in 2026, urging migration to these current IDs for production. Available via open-weights downloads or the official OpenAI-compatible API, they cater to B2B needs in operations, from financial modeling to software engineering agents. Open-Weights Access and Self-Hosting Options DeepSeek's commitment to openness shines in its R1 family, with full weights available under permissive licenses on Hugging Face and GitHub (github.com/deepseek-ai/DeepSeek-R1). Key open-weights models include (671B MoE), (long-context variant), and smaller distillations like (32B params on Qwen2.5

base) and (70B on Llama3). Self-hosting suits enterprises seeking data sovereignty and unlimited usage. Use frameworks like vLLM or TensorRT-LLM for inference, with quantization (e.g., 4-bit) reducing memory footprint— fits on 4x H100 GPUs post-quantization. Economics hinge on hardware: a 671B MoE like R1 requires 8-16x A100/H100 clusters for low-latency inference, amortized at $2-5/hour per GPU via cloud providers like AWS or self-owned data centers. Batch processing yields high throughput (e.g., 100+ tokens/sec), ideal for internal RAG pipelines, but upfront setup demands DevOps expertise. Official GitHub docs recommend temperature 0.5-0.7, no system prompts for reasoning, and self-hosting via Docker images for quick PoCs. Official DeepSeek API: Features and Model IDs For plug-and-play deployment, DeepSeek's API at platform.deepseek.com offers OpenAI-compatible endpoints with exact mo

del IDs: (successor to V3 for fast inference), (balanced reasoning), , and (full R1 access). Features include function calling, JSON mode, 128K-1M context (model-dependent), and streaming—perfect for agents and RAG. Pricing, per DeepSeek's official page (platform.deepseek.com/pricing) as of May 4, 2026, lists: - : $0.08 per 1M input tokens, $0.24 per 1M output tokens. - : $0.20 input / $0.60 output per 1M. - : $0.45 input / $1.80 output per 1M (higher for reasoning compute). These rates apply to pay-as-you-go tiers; volume discounts via enterprise plans reduce by 20-50%. No minimums, billed per token including reasoning chains. Secondary hosts like OpenRouter mirror these but add markups—always verify primaries. Math and Code Use Cases with Benchmarks DeepSeek R1 excels in math/code, per GitHub benchmarks: scores 85%+ on GSM8K (math reasoning), outperforming o1-mini, while full R1 hits 9

0%+ on MATH and HumanEval (code). V3 handles lighter tasks efficiently. Prompt best practices from GitHub: - Math: "Please reason step by step, and put your final answer within \boxed{}." Example: Solves AIME problems with chain-of-thought. - Code: Native support for Python/JavaScript generation, debugging via few-shot. Enterprise examples: Quantitative finance (derivative pricing), supply chain optimization (linear programming), or code agents for CI/CD. Benchmarks show MoE efficiency: R1 processes complex proofs 2x faster than dense peers at similar quality. Self-Host vs Hosted Economics Breakdown Compare via methodology: API costs scale with tokens (e.g., 1M queries at 10K ctx = $50 for as-of 2026-05-04), zero upfront. Self-hosting amortizes over volume—e.g., $10K/month H100 cluster handles 100M tokens/day for R1-Distill, breakeven at high scale ( 10B tokens/month). Factors: API wins

for variable loads/spikes; self-host for predictable, high-volume (e.g., 24/7 ops). Include inference multipliers: MoE activates subset, cutting costs 5x vs dense. Use calculators like DeepSeek's GitHub estimator or vLLM benchmarks. For 2026 enterprises, hybrid: self-host distillations, API for peak