DeepSeek V4 Release Date Predictions
The model has already been delayed (more on this later), but as of the time of writing, multiple reliable signals independently point to a late April to mid-May 2026 window, making it highly likely:
- According to Reuters. On 3 April 2026, Reuters reported that V4 is likely to launch within the next few weeks. Taking this literally implies a window ending around early May.
- According to Polymarket. As of 18 April 2026, the Polymarket DeepSeek V4 released by…? market assigns a 79% probability of release by 30 April and 92% by 15 May. Polymarket usually ends up within about 5–10 percentage points of what actually happens on AI release dates, so it’s a reliable signal to watch.
- According to the historical release pattern of DeepSeek models. DeepSeek R1 launched on 20 January 2025, immediately before the Lunar New Year holiday, and the team does not typically ship major models during Chinese national holidays. The next major window after the 17 February 2026 Lunar New Year is the post-Labour Day period in early May, which is consistent with the other signals.
Our best estimate, based on all of the above: DeepSeek V4 will launch in the last two weeks of April or the first two weeks of May 2026.
DeepSeek V4 Release Date Timeline
This section tracks every credible public signal in the V4 release cycle, in reverse chronological order.
18 April 2026. Polymarket shows 79% odds of release by 30 April and 92% by 15 May. The earlier Released by March 31 market resolved to No on 31 March.
16 April 2026. Nvidia CEO Jensen Huang commented on the Dwarkesh Podcast that DeepSeek optimising V4 for Huawei chips instead of American hardware would be "a horrible outcome" for the United States.
3 April 2026. Reuters reported that V4 is likely to launch in the next few weeks and will run on Huawei's Ascend 950PR processor for inference. The same report confirmed DeepSeek has been working with Huawei and Cambricon for months to rewrite parts of the model stack.
1 April 2026. Alleged V4 benchmarks leaked on X via account @bridgemindai, claiming 83.7% on SWE-Bench Verified. Community notes flagged parts of the leak as likely fabricated (an AIME 2026 score of 99.4% is mathematically impossible under the official scoring system). But the SWE-Bench number has since reappeared in another leak, making it plausible.
31 March 2026. Polymarket DeepSeek V4 Released by March 31 market resolved to No.
18 March 2026. Reuters confirmed that Hunter Alpha, a stealth model that appeared on OpenRouter and generated V4-related speculation, was actually Xiaomi's MiMo-V2-Pro rather than DeepSeek V4. The same report noted Chinese media pointing to an April launch window.
9 March 2026. An unannounced V4 Lite (also referred to as Sealion-lite) appeared quietly on the DeepSeek platform. Developers testing it notice a 30% inference speed improvement over V3.2 and a 94% context recall at 128K tokens.
26 February 2026. DeepSeek denied Nvidia and AMD early access to V4 for optimization, while domestic suppliers including Huawei did get early access.
February 2026. No launch as the model was held back for further testing.
9 January 2026. DeepSeek planned to launch a coding-focused AI model in February with an emphasis on code generation and long coding prompts. This was the first credible public signal of the V4 cycle.
Why Has the DeepSeek V4 Release Been Delayed?
Hardware migration. DeepSeek has spent several months rewriting its stack to run on Huawei's Ascend 950PR and Cambricon chips instead of Nvidia. This is the first frontier-class AI model explicitly optimised away from CUDA, and the engineering work to validate inference at trillion-parameter scale on new hardware has taken longer than originally planned.
Training stability at 1T parameters. Trillion-parameter MoE models are notoriously hard to train without gradient instability. DeepSeek's own Manifold-Constrained Hyper-Connections (mHC) paper from January 2026 addresses this specific problem, and the delay suggests the team wanted to fully validate the architecture before release.
V4 Lite testing. The quiet release of V4 Lite in early March functions as a public stress test of the V4 architecture at a smaller scale. A full flagship release typically follows this kind of validation by a few weeks.
What to Watch for Launch
If you want to be the first to know when DeepSeek V4 launches, here are the sources that will show it first.
- DeepSeek API documentation. A new deepseek-v4 or similarly named model ID will appear in official API docs.
- Hugging Face. DeepSeek has open-sourced every previous generation (V2, V3, V3.2, R1) on Hugging Face shortly after release. Watch the DeepSeek organization page for V4 weights.
- DeepSeek GitHub. Architecture papers, inference code, and technical reports are usually posted within hours of model release on the model’s Github.
- Overchat AI waitlist. If you'd rather have the model pushed to you, join the waitlist on Overchat AI and we'll notify you the moment V4 is available on our platform.
FAQ
Has DeepSeek V4 launched?
No. As of now, DeepSeek V4 has not launched yet. The DeepSeek API still serves V3.2 models, and no official announcement has been made.
When will DeepSeek V4 be released?
The most likely window is late April to mid-May 2026.
Why has DeepSeek V4 been delayed?
Two main reasons. First, DeepSeek rewrote parts of the model stack to run inference on Huawei's Ascend 950PR chips instead of Nvidia, which required months of additional engineering. Second, trillion-parameter MoE models are hard to train reliably, and the team appears to have wanted to fully validate the architecture through the V4 Lite release before the flagship launch.
Where will I be able to use DeepSeek V4 first?
Overchat AI will add DeepSeek V4 support within hours of official release. Subscribe to updates to be notified as soon as it drops.
Will DeepSeek V4 be open-source?
DeepSeek has released every previous flagship (V2, V3, V3.2, R1) as open-weight under a permissive license, and V4 is expected to follow the same pattern, likely under Apache 2.0.
What happens to this page after DeepSeek V4 launches?
This page will be updated with the confirmed launch date, benchmarks, and access instructions. The waitlist signup will be replaced with a direct link to use V4 on Overchat AI.
Bottom Line
Although DeepSeek V4 has not yet been launched, every credible signal points to a release within weeks. The model is shaping up to be the most powerful open-source AI model in history and, according to unverified leaks, it may even be the world's best AI code model. To learn more, check out our full breakdown of DeepSeek V4, based on everything we currently know about the model.
Key Takeaways
- DeepSeek V4 has not launched yet, and the official API still serves V3.2.
- The launch is expected sometime in late April to mid-May 2026, as per Reuters.
- Polymarket gives a 79% probability that the model will come out by 30 April.
- DeepSeek rewrote its stack to run on Huawei Ascend 950PR instead of Nvidia, which took months of extra engineering — this is the reason why the model was delayed.
