Is DeepSeek V4 Released Yet? Official Status, Leaked Specs, and What to Use Now

DeepSeek V4 has not been officially released. As of April 2026, there is no V4 model available on the DeepSeek API, app, or website. What you are seeing in search results and social media is a mix of credible reporting from Reuters, third-party speculation, and SEO-bait articles presenting leaked specs as confirmed facts. This article gives you the verified status, what the credible rumors actually say, and what you should be running today while you wait.

Is DeepSeek V4 Released?

No. As of April 11, 2026, DeepSeek has not published a V4 model ID, a pricing page, a technical report, or any announcement on their official channels. The DeepSeek API changelog still lists DeepSeek-V3.2 as the current production model. The official deepseek-chat and deepseek-reasoner API endpoints map to V3.2 and V3.2's thinking mode respectively.

The confusion stems from two sources: Reuters reported on April 4, 2026 (citing The Information) that DeepSeek V4 would likely launch "within the next few weeks." That is a credible signal but not a launch announcement. Earlier in the year, V4 was expected around mid-February 2026 to coincide with Lunar New Year — that window passed without a release.

Why the Release Date Has Slipped Multiple Times

DeepSeek V4 is reported to be the first frontier-class AI model built to run on Chinese semiconductor hardware — specifically Huawei's Ascend 950PR chips. That is a significantly harder engineering challenge than iterating on an existing NVIDIA-based stack. DeepSeek and Huawei teams have reportedly been rewriting core inference infrastructure to adapt to a completely different chip architecture. Chip-level optimization at this scale introduces delays that are difficult to predict from the outside.

A secondary factor: DeepSeek appears to be staging a proper capacity build-out before launch rather than releasing to a waitlist. The V3 and V3.2 launches both resulted in API overload. A model running on Huawei Ascend at trillion-parameter scale requires careful capacity planning before general availability.

How to Track the Official V4 Launch

The most reliable sources for a deepseek v4 release date confirmation, in order of trustworthiness:

• DeepSeek API Docs changelog — api-docs.deepseek.com/updates — every major model update has been posted here first
• DeepSeek on X (@deepseek_ai) — official launch announcements appear here simultaneously with the docs
• Hugging Face (deepseek-ai) — model weights are posted here if V4 ships as open-weights (V3 and V3.2 were both open-weights)
• DeepSeek GitHub — technical reports and model cards accompany every major release

Do not trust third-party "release date trackers" or countdown sites — they are SEO plays, not official sources.

DeepSeek V4 Release Date Rumors and Leaked Specs

Current credible reporting points to a late April 2026 launch. Reuters (April 4) cited people familiar with the matter. Polymarket prediction markets had V4 releasing before March 31 at roughly 40% — that window expired, shifting current expectations to late April or early May 2026.

The Huawei Ascend 950PR Story

The most significant confirmed detail about deepseek v4 is not its benchmark scores — it is the chip infrastructure. Reuters confirmed that V4 will run entirely on Huawei's Ascend 950PR chips, making it the first frontier-class AI model trained and served on Chinese semiconductor hardware. This matters for two reasons:

• Geopolitical independence: Huawei chips are not subject to US export controls the way NVIDIA H100s and A100s are. If V4 performs competitively on Ascend hardware, it validates China's AI hardware supply chain as a credible alternative.
• Performance implications: Ascend 950PR is a less mature inference platform than NVIDIA's current stack. DeepSeek has reportedly invested heavily in custom kernels for Huawei hardware. How well this translates to throughput and latency in production is unknown until launch.

Rumored V4 Specs (Unverified)

The following specs appear in multiple third-party reports but have not been confirmed by DeepSeek. Treat them as directional signals, not architectural facts:

• Parameter count: ~1 trillion total (likely MoE with a much smaller active parameter count per token) — unverified
• Context window: 1 million tokens (up from 128K in V3.2) — unverified
• Inference speed: 1.8x faster than V3.2 — unverified, source is a single leak
• Pricing: Near-Claude-Opus-level quality at roughly 1/50th the input cost — unverified, but consistent with DeepSeek's historical pricing trajectory
• Multimodal support: Vision input rumored — unverified

None of these specs appear in DeepSeek's official API documentation or any technical report as of April 2026. Do not make infrastructure decisions based on unconfirmed V4 capability claims.

What to Use While You Wait: DeepSeek V3.2 Today

While the industry speculates about deepseek v4, DeepSeek V3.2 is a genuinely capable production model available right now. It is the official successor to V3.2-Exp, and both deepseek-chat and deepseek-reasoner API endpoints were upgraded to V3.2 on launch. For context on how V3.2-Exp evolved, see our DeepSeek V3.2-Exp API guide.

Key V3.2 capabilities that are production-ready today:

• Thinking integrated into tool-use: V3.2 is the first DeepSeek model to run the reasoning chain inside a tool-call loop. This is significant for agentic workflows where you want the model to reason before deciding which tool to invoke.
• 1,800+ environment agent training: V3.2 was trained on 85,000+ complex instructions across 1,800+ simulated environments. In practice: noticeably better multi-step instruction-following compared to V3.
• 128K context window: Sufficient for most codebase summarization, long document analysis, and multi-file refactoring tasks.
• OpenAI-compatible API: Drop-in replacement for most OpenAI SDK integrations — swap the base URL and API key, keep your existing code.

V3.2 vs V3.2-Speciale: Which Should You Use?

DeepSeek released two V3.2 variants simultaneously:

• DeepSeek-V3.2: General-purpose model. Available via API, web app, and mobile. Supports tool-use in both thinking and non-thinking modes. This is your production model for most use cases.
• DeepSeek-V3.2-Speciale: Competition-level math and algorithmic reasoning — gold-level at IMO, ICPC World Finals, and IOI 2025. Currently API-only, no tool-use support. Use this for pure reasoning tasks where you do not need agent capabilities. For full setup details, see our DeepSeek V3.2-Speciale installation guide.

Accessing DeepSeek V3.2 via API

If you are already using OpenAI's Python SDK, the migration is a single-line change:

from openai import OpenAI

client = OpenAI(
    api_key="your_deepseek_api_key",
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    model="deepseek-chat",
    messages=[
        {"role": "user", "content": "Explain MoE architecture in one paragraph."}
    ]
)
print(response.choices[0].message.content)

For V3.2-Speciale (reasoning mode), use model="deepseek-reasoner". One important note for budget planning: deepseek-reasoner bills thinking tokens separately from output tokens. On complex reasoning tasks the thinking token count can be 2-5x the output token count — factor this into your cost projections.

DeepSeek V3.2 vs Alternatives: Benchmarks and Pricing (April 2026)

If you are deciding whether to use DeepSeek V3.2 now or hold for V4, here is how the current landscape compares. For a forward-looking architectural comparison of how V3 capabilities evolved toward what V4 promises, see our DeepSeek V3 vs V4 deep dive.

• DeepSeek V3.2: ~$0.27/M input | 128K context | open weights | competitive coding performance
• Claude Opus 4.6: ~$15/M input | 200K context | closed | approximately 82% SWE-bench Verified — best for hard multi-file coding
• GPT-5.4: ~$10/M input | 128K context | closed | approximately 80% SWE-bench Verified — broadest multimodal capabilities
• Gemini 3.1 Pro: ~$2/M input | 1M context | closed | 80.6% SWE-bench Verified — best price/performance for long-context tasks
• Qwen 3.5: ~$0.40/M input | 128K context | open weights | optimized for edge and local deployment
• MiniMax M2.5: ~$0.75/M input | 1M context | partial open | approximately 80.2% SWE-bench Verified

SWE-bench Verified scores marked with "approximately" are based on third-party reporting and have not been confirmed in official DeepSeek V3.2 benchmarks.

The cost efficiency case for DeepSeek V3.2 is clear: for high-volume code generation, data transformation, or classification tasks, ~$0.27/M input is roughly 1/55th the cost of Claude Opus 4.6 at competitive output quality. If your workload is latency-sensitive and quality-critical, Opus or GPT-5.4 remain the safer choice at higher cost.

Wait for V4 or Ship With V3.2 Now?

The practical decision comes down to your timeline and requirements:

• Shipping in the next 4-8 weeks: Use V3.2. It is production-stable, the API has been reliable since launch, and V4's real-world API performance will not be known for weeks after release. Do not block a production launch on an unconfirmed model.
• Need more than 128K context today: V3.2 tops out at 128K. Use Gemini 3.1 Pro or MiniMax M2.5 for long-context tasks — both offer 1M context windows at reasonable prices.
• Require open weights: V3.2 weights are available on Hugging Face. V4's open-weights status is unconfirmed — the Huawei chip architecture makes this especially uncertain.
• Designing new AI-native infrastructure: A 4-6 week wait to see V4's actual API specs is reasonable, as long as you are not blocking existing development on that wait. Build on V3.2 and migrate if V4 delivers the rumored context window and performance gains.

The best engineers do not wait for the next model — they build with what is production-stable today and upgrade when the new model is proven. DeepSeek V3.2 is proven. DeepSeek V4 is a credible rumor with a Reuters source behind it.

When V4 does launch, run your own benchmark suite on your specific task distribution before migrating. Aggregate benchmark scores do not tell you how the model performs on your particular workload. Test with your own inputs, your own success criteria, and your own throughput requirements before switching production traffic.