DeepSeek V3 launched in late 2024 as the company's flagship general-purpose model. It didn't stay still. Through 2025 and into 2026, V3 went through multiple major revisions before eventually evolving into V4. If you last checked DeepSeek six months ago, a lot has changed. Here's what's new across the V3 line and what the deepseek v4 update means.
From V3 to V3.2: The Incremental Leaps
The original DeepSeek V3 arrived in December 2024 as a 671-billion-parameter MoE model with 37 billion active parameters. It was immediately notable for two things: competitive benchmark scores against GPT-4 and a training cost of roughly $5.6 million, a fraction of what competitors spent.
V3.1 arrived in mid-2025 with improvements to language consistency. The model had a known issue with mixing Chinese and English mid-response, and V3.1 largely fixed it. Agent capabilities got attention too, with better tool calling and code agent performance.
V3.2 (September 2025) was the biggest pre-V4 update. The key changes:
- Thinking mode merged into the main model. Before V3.2, the reasoning model (R1) and the chat model (V3) were separate. V3.2 unified them. The standard mode handled quick tasks, and switching to thinking mode gave you DeepThink reasoning without changing the underlying model.
- Context window expanded to 128K tokens. Up from the original 64K, this meant the model could process roughly 200 pages of text in a single session.
- Tool calling got tighter. V3.2 improved the model's ability to use external tools like web search and code execution without losing track of the conversation.
DeepSeek API V3: What Developers Got
The DeepSeek API mirrored the chat product's evolution. Key API-specific changes during the V3 era:
September 2025: Both deepseek-chat and deepseek-reasoner endpoints were upgraded to V3.2. The legacy separation between chat and reasoning models went away. Developers could toggle thinking mode via a parameter instead of routing to different endpoints.
December 2025: V3.2-Speciale launched as a temporary experimental endpoint with tool-free capabilities, testing features destined for V4. It expired after two weeks, which was clearly a preview of things to come.
Pricing through the V3 era: V3 API pricing stayed at roughly $0.27/million input tokens and $1.10/million output tokens, making it one of the cheapest frontier-class APIs available. For comparison, GPT-4o at the time was charging $2.50/million input and $10/million output.
V3 → V4: The Big One
April 24, 2026 marked the end of the V3 era. DeepSeek V4 Preview launched with two models replacing the V3 line:
V4-Pro: 1.6 trillion total parameters, 49 billion active. A massive jump from V3's 671B total and 37B active.
V4-Flash: 284 billion total parameters, 13 billion active. Smaller than V3 in total size but more efficient.
Both V4 models shipped with:
- 1-million-token context window as standard. That's roughly 750,000 words, enough for three full-length novels.
- 73% lower inference compute compared to V3.2 at equivalent context lengths.
- 90% smaller KV cache, making long-context processing dramatically cheaper.
- Native image understanding added to the model itself.
The API inherited both models. deepseek-v4-pro and deepseek-v4-flash replaced the old deepseek-chat and deepseek-reasoner endpoints, though the old names still work until July 2026 as aliases pointing to V4-Flash.
V4 Formal Release: July 2026
The July 2026 general availability release of V4 brought the Flash model to production-grade stability. Improvements over the April preview:
- Agent performance: Task completion rates improved by roughly 21% in V4-Flash formal compared to preview, based on internal benchmarks.
- Industry-first peak/off-peak pricing for API users. Cheaper rates during low-traffic hours.
- OpenAI Responses API compatibility. V4-Flash can now plug directly into ChatGPT Codex as a model provider, letting developers use DeepSeek's cheaper inference inside OpenAI's coding agent.
- Domestic chip support finalized. V4 runs on Huawei Ascend processors with 85% compute utilization, closing the loop on the company's hardware independence strategy.
What This Means for Users
If you're using the free DeepSeek app, you're already on V4. The model upgrade happened server-side. You didn't need to update anything. Your chat interface now runs on a model with a million-token context window, image understanding, and better agent capabilities than what you had in early 2026.
For developers on the API, the old deepseek-chat and deepseek-reasoner endpoints will stop working on July 24, 2026. If you haven't migrated to deepseek-v4-pro or deepseek-v4-flash, now's the time. The new pricing is competitive: V4-Flash input at $0.28/million tokens, output at $1.10/million tokens.
The V3 era was defined by proving that a small team with limited hardware could compete with billion-dollar labs. The V4 era, judging by the first few months, is about making that capability accessible to everyone.