DeepSeek V4 Pro and Flash: 1M context for everyone, now on Napsix
DeepSeek released V4 Pro (1.6T params) and V4 Flash (284B): both open-source, with 1M context, compatible with the OpenAI and Anthropic APIs. Legacy aliases retire today. Already wired into the AI Gateway.
DeepSeek released its V4 series (Pro + Flash) on April 24, 2026 — both models open-source and shipping with a 1-million-token context window — and today, July 24, the legacy deepseek-chat/deepseek-reasoner aliases are permanently retired. Both models are already wired into Napsix's AI Gateway.
Two sizes, one context standard
- DeepSeek-V4-Pro: 1.6 trillion total parameters, 49 billion activated. Rivals the world's top closed-source models on agentic coding (open-source state-of-the-art), leads all open models on general knowledge (trailing only Gemini 3.1 Pro), and beats all current open models on math/STEM/coding, rivaling top closed-source models.
- DeepSeek-V4-Flash: 284 billion total parameters, 13 billion activated. Reasoning closely approaches V4-Pro, performs on par on simple agent tasks, with faster responses and the series' lowest price — ideal for high-volume traffic.
Structural innovation: 1M context without the usual cost
Both models share the same novel attention mechanism — token-wise compression combined with DSA (DeepSeek Sparse Attention) — making 1M context the default standard across all official DeepSeek services, at drastically lower compute and memory cost than prior generations (DeepSeek-V3.2 and R1 topped out at 128K-164K).
Dedicated optimizations for agents
DeepSeek-V4 is seamlessly integrated with leading agents like Claude Code, OpenClaw, and OpenCode, and it's already driving DeepSeek's own in-house agentic coding. The image below shows a real example: a PDF generated end to end by DeepSeek-V4-Pro.
API compatibility
DeepSeek's API exposes both models under the same base URL, compatible with OpenAI's ChatCompletions interface and with Anthropic's API — just change the model name to deepseek-v4-pro or deepseek-v4-flash. Both support interchangeable thinking and non-thinking modes per request.
Pricing
| Model | Input (cache miss) | Input (cache hit) | Output |
|---|---|---|---|
| V4-Pro | $0.435/M | $0.0036/M | $0.87/M |
| V4-Flash | $0.14/M | $0.0028/M | $0.28/M |
(The 75%-off promo on V4-Pro became the official price as of May 31, 2026 — not a temporary discount. The original list price was $1.74/$3.48 per million.)
Deprecation notice
The legacy deepseek-chat and deepseek-reasoner aliases (which routed to V4-Flash's non-thinking/thinking modes) retire today, July 24, 2026 at 15:59 UTC. Any integration still using those names needs to migrate to deepseek-v4-pro / deepseek-v4-flash — the base_url doesn't change, only the model parameter.
Open-weight
All 4 checkpoints (V4-Pro, V4-Flash, and their base versions) are available on Hugging Face under an open license, with the full technical report published the same day as the release.
On Napsix
DeepSeek V4 Pro and Flash are already active in Napsix's model catalog (served via OpenRouter). Pick them from the model selector in any XIA chat for deep reasoning work (Pro) or high-volume, low-cost tasks (Flash).
