DeepSeek official models list 2026
Last updated: April 24, 2026
This is a canonical lineup page built for users who search terms like "deepseek official models list 2026" or "deepseek official models list v4". Instead of only listing names, the page adds routing context so visitors immediately understand which models are available in this site, where each model runs today, and what task profile each model is optimized for. That extra context helps both human readers and AI answer engines extract a reliable short answer without mixing old model aliases with current production names.
Verification follows a strict order: first DeepSeek API Docs changelog and pricing pages, then OpenRouter model pages for `deepseek/deepseek-v4-flash` and `deepseek/deepseek-v4-pro`, then Replicate model cards for V3.1, R1, Math-7B Instruct, Janus-Pro-7B, and VL2. When labels differ between platforms, we keep official model names in this list and treat provider-specific naming as routing metadata. This prevents keyword cannibalization between your `/models`, `/news`, and `/official-sources` pages while still giving users direct verification paths.
The DeepSeek official models list 2026 includes V4, V4 Flash, V4 Pro, V3.1, R1, Math-7B Instruct, Janus-Pro-7B, and VL2. Current routing in this site uses OpenRouter for V4 Flash and V4 Pro, while legacy and specialist models are routed through Replicate.
| Model | Category | Best-fit workload | Route | Status |
|---|---|---|---|---|
| DeepSeek V4 | Multimodal | Flagship multimodal workloads, High-complexity reasoning | OpenRouter route | available |
| DeepSeek V4 Flash | Text | Realtime chat, High-throughput automation | OpenRouter route | available |
| DeepSeek V4 Pro | Reasoning | Complex reasoning, Code and architecture reviews | OpenRouter route | available |
| DeepSeek V3.1 | Text | General chat, Code generation | Replicate route | available |
| DeepSeek R1 | Reasoning | Multi-step reasoning, Math and proofs | Replicate route | available |
| DeepSeek Math-7B | Math | Math tutoring, Symbolic reasoning | Replicate route | available |
| DeepSeek Janus-Pro-7B | Multimodal | Text-to-image, Visual reasoning | Replicate route | available |
| DeepSeek VL2 | Multimodal | OCR, Document analysis | Replicate route | available |
Keep this list as your canonical entry point for model naming. In product copy, avoid mixing deprecated aliases like `deepseek-chat` and `deepseek-reasoner` with new model IDs once migration is complete. In engineering docs, map each model to its current provider route and note that provider ceilings can differ by account tier, region, or queue pressure.
For user experience, route first-pass prompts to V4 Flash and escalate only selected prompts to V4 Pro. This keeps response speed consistent and preserves high-quality reasoning for complex tasks such as architecture analysis, long-horizon planning, and full repo troubleshooting. The result is a cleaner product story around V4 while legacy models remain available for niche workloads like OCR-heavy pipelines or math-specific evaluation.