TLDR
- DeepSeek V4 is the next flagship model from Chinese AI lab DeepSeek, expected to be the most capable open-source AI model ever released.
- Hybrid architecture. V4 combines reasoning and non-reasoning in a single model, retiring the split between V3.X and R1 and likely ending the R2 line entirely.
- 1 trillion parameters. Approximately 37B active per token via Mixture-of-Experts, keeping inference costs close to a mid-sized model.
- 1M+ token context window, enabled by DeepSeek Sparse Attention (DSA). For comparison, Claude Opus 4.7 is 200K.
- New memory and training methods. Engram conditional memory for near-infinite context recall, plus Manifold-Constrained Hyper-Connections (mHC) for stable trillion-parameter training.
- First frontier model on Chinese silicon. Inference will run on Huawei's Ascend 950PR via the CUDA-compatible CANN Next framework, the first major AI model explicitly optimised away from Nvidia.
- Leaked 83.7% on SWE-Bench Verified — unverified, but if accurate, V4 would be the best coding model in the world.
- Open-source release under a permissive Apache 2.0 license is expected, with weights available via Hugging Face.
- On Overchat AI. You can use DeepSeek V3.2 today, and V4 will be added as soon as it's available. Join the waitlist to be notified when V4 launches.
Latest Updates
16 April 2026. Nvidia CEO Jensen Huang warned on the Dwarkesh Podcast that DeepSeek optimising V4 to run on Huawei chips instead of American hardware would be "a horrible outcome" for the United States. The comment confirms how strategically important this release has become.
4 April 2026. Reuters confirmed — citing The Information — that V4 will run inference on Huawei's new Ascend 950PR processor. Training reportedly still happened on Nvidia Blackwell chips. Alibaba, ByteDance, and Tencent have placed bulk orders for hundreds of thousands of Ascend 950PR chips in preparation, and chip prices jumped roughly 20% in weeks.
1 April 2026. Alleged V4 benchmarks leaked on X via account @bridgemindai, claiming 83.7% on SWE-Bench Verified, 88.4% on IMO Answer Bench, and 23.5% on FrontierMath Tier 4. Community notes flagged an AIME 2026 score of 99.4% as impossible under the official scoring system, so treat the full leak as unverified and likely partially fabricated. The SWE-Bench number has reappeared across several secondary sources and is the most plausible single data point.
9 March 2026. An unannounced "V4 Lite" (also referred to as Sealion-lite, ~200B parameters) appeared quietly on the DeepSeek platform. Developers testing it reported a 30% inference speed improvement over V3.2 and 94% context recall at 128K tokens. This effectively validates the core V4 architecture ahead of the full flagship launch.
Join the waitlist on Overchat AI to be notified the moment V4 launches.
What Is DeepSeek 4?
About DeepSeek V4. DeepSeek V4 is the first whole-numbered update from the Hangzhou-based AI company since DeepSeek V3 was released in December 2024. It is the next flagship model, used by almost one million people per month on the official chatbot platform alone.
The model. DeepSeek has built a reputation as a highly capable coding model optimised for reasoning and complex tasks. V4 continues this heritage — primarily a coding and programming model, it will likely excel in logic and practical work rather than AI writing.
Hybrid architecture. V4 is a hybrid model that supports both reasoning and non-reasoning tasks. That means the distinction between R1 and V3.X is going away, and DeepSeek R2 likely isn't coming at all.
Scale. Leaked figures put V4 at approximately 1 trillion total parameters, with roughly 37 billion active per token via the Mixture-of-Experts architecture. That keeps inference costs close to a 37B model while giving V4 the capacity of a trillion-parameter system.
DeepSeek V4 Main Features
Manifold-Constrained Hyper-Connections (mHC). On 1 January 2026, DeepSeek published a research paper co-authored by founder Liang Wenfeng describing a new training method. The approach changes how information flows through the model's layers and enables better performance without significant additional computational cost. It also addresses gradient instability at trillion-parameter scale, which has historically made large MoE models unreliable to train.
Engram Memory System. Published on 13 January 2026, this conditional memory technology enables the model to selectively retain and recall information based on context. In practice, the model consumes context much more slowly and stays accurate over longer documents, codebases, and chats. Early V4 Lite tests report 94% context recall at 128K tokens, compared to roughly 45% for V3.2 at the same length.
DeepSeek Sparse Attention (DSA). Enables context windows to exceed one million tokens. For comparison, Claude Opus 4.7 has a 200,000-token window. This is one of the largest context windows on any publicly available model.
Mixture-of-Experts (MoE) Architecture. Continuing from V3, the model activates only a fraction of its total parameters per request, dramatically improving efficiency.
Huawei Ascend 950PR hardware path. V4 is the first frontier AI model explicitly optimised for Chinese domestic silicon. Inference will run on Huawei's Ascend 950PR via the CANN Next heterogeneous computing framework, which is CUDA-compatible. This matters strategically because it is the first frontier model proving that a competitive AI system can run without Nvidia hardware in production.
Who Created DeepSeek?
DeepSeek was founded in July 2023 by Liang Wenfeng, a 40-year-old entrepreneur who also co-founded the quantitative hedge fund High-Flyer.
Liang studied at Zhejiang University, earning degrees in electronic information engineering and communication engineering. In 2015, he co-founded High-Flyer with classmates. The hedge fund relies on AI and mathematical models for trading decisions and exceeded 100 billion yuan (roughly $14 billion) in assets by 2021.
DeepSeek operates as a subsidiary of High-Flyer, which remains its sole financial backer. Liang holds an 84% stake. The company takes an unusual hiring approach: recruiting primarily from Chinese universities, often bringing on recent graduates, and seeking talent from non-computer-science backgrounds including poetry and advanced mathematics.
Liang has stated that the company's goal isn't quick profits but advancing AI toward artificial general intelligence.
How Good Will DeepSeek V4 Be?
The leaked benchmarks. On 1 April 2026, purported V4 benchmarks surfaced on X claiming the following numbers on SWE-Bench Verified:
- DeepSeek V4: 83.7%
- GPT-5.2 High: 80.0%
- Kimi K2.5 Thinking: 76.8%
- Gemini 3.0 Pro: 76.2%
- DeepSeek V3.2 Thinking: 73.1%
If accurate, this would make V4 the best coding model in the world. However, the same leak also claimed an AIME 2026 score of 99.4%, which is mathematically impossible under the official scoring system (the max achievable is 99.2% or 100%). Community notes flagged the image as likely fake. Treat the 83.7% SWE-Bench number as an unverified claim, not a confirmed benchmark.
Context on internal benchmarks. Sources who spoke to The Information said V4 outperforms Claude and GPT models in DeepSeek's internal tests. Internal tests should always be treated with caution, but DeepSeek has a track record of underplaying rather than exaggerating its releases. When DeepSeek R1 launched in January 2025, it matched OpenAI's models on maths and reasoning while reportedly costing just $6 million to train.
Read more about DeepSeek vs ChatGPT.
When Is the DeepSeek 4 Release Date?
DeepSeek is targeting a release around mid-February 2026, which is likely to coincide with the Lunar New Year celebrations on 17 February.
If true, this would be the same timing as that of the release of DeepSeek's R1 on 20 January 2025, just before China's week-long Lunar New Year holiday. This release triggered a $1 trillion sell-off in tech stocks, including $600 billion from Nvidia alone.
However, DeepSeek has declined to officially confirm the release date.
Where to Access DeepSeek V4
Overchat AI
Overchat AI is an all-in-one platform that provides access to multiple AI models including DeepSeek. You can currently access DeepSeek V3.2 online or via mobile apps for iOS and Android, with a free tier of 50 messages per month. V4 will be added to Overchat AI as soon as it becomes available. Join the waitlist to be notified when V4 launches.
Official DeepSeek platforms
- DeepSeek Chat: free web-based chatbot
- DeepSeek API: pay-per-use pricing, significantly cheaper than competitors
- Mobile apps for iOS and Android
Open-source access
DeepSeek is expected to release V4 as an open-weight model under a permissive Apache 2.0 license. You will be able to download weights from Hugging Face, run locally using tools like Ollama or vLLM, self-host on cloud platforms, and fine-tune for specific use cases. Because of MoE efficiency and INT8/INT4 quantization, V4 should run on consumer hardware such as dual RTX 4090s or a single RTX 5090.
What Are the Best DeepSeek V4 Alternatives?
Here are the AI models V4 is expected to compete with in benchmarks:
- Claude Opus 4.7: Anthropic's latest flagship. Known for strong reasoning and long-context processing.
- GPT-5.5: OpenAI's current flagship, widely used for writing, coding, and analysis across consumer and enterprise applications.
- Gemini 3.0: Google's latest multimodal model with deep ecosystem integration.
- Kimi K2.5: Moonshot AI's flagship, the closest Chinese competitor to V4 on coding benchmarks.
FAQ
Is DeepSeek V4 free to use?
Yes — the chatbot interface will likely remain free. API access is pay-per-use but significantly cheaper than competitors. If released as open-source, you can run it locally for free as well.
Will DeepSeek V4 work outside China?
Yes. Based on DeepSeek's established pattern with V3 and V3.2, international API access through api.deepseek.com is expected to continue. This has not been officially confirmed for V4, and the geopolitical context around US–China chip exports is worth monitoring.
Is DeepSeek safe to use?
Yes. You may have heard that DeepSeek's app has been banned in some countries because of privacy concerns. This is mostly political, driven by the company's Chinese origin. DeepSeek is used by over 1 million people monthly, and many users say it's the best AI model available today.
Bottom Line
If the rumors are true, DeepSeek V4 will be the best AI coding model yet. Here's a brief summary of the most important things you need to know about the upcoming release.
Key Takeaways
- DeepSeek V4 is the next flagship open-source AI model from Chinese AI lab DeepSeek, expected to be the most capable open-source model ever released.
- Architectural leap. Hybrid reasoning, 1 trillion total parameters (~37B active), 1M+ token context, and new training methods (mHC, Engram, DSA).
- Huawei Ascend 950PR. V4 is the first frontier AI model running production inference on Chinese silicon, a major step away from Nvidia.
- Leaked 83.7% on SWE-Bench Verified. Unverified, but consistent with DeepSeek's track record of quiet releases that outperform expectations.
- On Overchat AI. DeepSeek V3.2 is available today, and V4 will be added the moment it launches. Join the waitlist.
For more on the timing, read our full article on the DeepSeek V4 release date.
