A year ago, DeepSeek's R1 model made Silicon Valley collectively spit out its coffee. A Chinese startup, training competitive models for under $6 million while U.S. labs burned through billions? That felt impossible. Well, DeepSeek just did it again. On April 24, 2026, the company released a preview of DeepSeek V4, and the numbers are genuinely staggering.
Two models shipped today: DeepSeek-V4-Pro with 1.6 trillion parameters, and DeepSeek-V4-Flash at 284 billion parameters. Both are open source. Both feature a 1 million token context window. And both are designed to run on Huawei's Ascend AI chips — not just Nvidia hardware.
Here's what actually matters for developers and teams evaluating this model.
Key Takeaways
- DeepSeek V4 preview launched April 24, 2026 in two variants: V4-Pro (1.6T parameters) and V4-Flash (284B parameters).
- Both models are open source, continuing DeepSeek's track record of releasing model weights for developers to download, run locally, and modify.
- 1 million token context window — a massive jump from the 128K context of the previous V3 model, enabling repo-scale coding and massive document processing.
- V4-Pro leads all open-source models in math and coding benchmarks, trailing only Google's closed-source Gemini 3.1-Pro in world knowledge tasks.
- Optimized for Huawei's Ascend AI chips, with Huawei announcing "full support" on launch day. Cambricon Technologies also confirmed compatibility.
- Agent-focused capabilities — V4 is optimized for popular agentic tools like Anthropic's Claude Code and OpenClaw.
- V4-Flash pricing matches DeepSeek V2 from June 2024, making it one of the cheapest frontier-class models available anywhere.
- Not a production-ready release yet — this is a preview. Teams should evaluate, not deploy to production.
What Actually Shipped: DeepSeek V4-Pro and V4-Flash
DeepSeek didn't release one model. They released two, targeting very different use cases.
V4-Pro is the flagship. At 1.6 trillion parameters, it's DeepSeek's biggest model ever by that metric. According to the company's own benchmarks, it beats every other open-source model in math and coding tasks. For world knowledge, it trails only Google's Gemini 3.1-Pro — a closed-source model that costs significantly more to access.
DeepSeek described V4-Pro's performance as "marginally short" of OpenAI's GPT-5.4 and Gemini 3.1-Pro. They estimate the developmental gap at roughly 3 to 6 months behind the absolute frontier. For an open-source offering, that's remarkably close.
V4-Flash is the lighter, faster sibling. At 284 billion parameters, it keeps similar reasoning abilities to the Pro version but runs cheaper and responds quicker. Here's the kicker: its token pricing matches DeepSeek V2 from mid-2024. That arguably makes it the cheapest cutting-edge model you can use right now. The South China Morning Post confirmed this pricing detail, noting V4-Flash as "one of the cheapest cutting-edge models available on the market."
Both models share one critical new feature: a 1 million token context window. The previous V3 topped out at 128,000 tokens. Going from 128K to 1M isn't incremental. It fundamentally changes what you can do. Processing entire codebases, long legal documents, or multi-chapter research papers in a single pass becomes practical rather than theoretical.
The Hardware Story: Huawei Chips Take Center Stage
This is where things get geopolitically interesting.
Within hours of the DeepSeek V4 preview going live, Huawei announced "full support" across its Ascend chip lineup and supernode systems. Cambricon Technologies also quickly confirmed compatibility. Reuters had reported earlier in April that DeepSeek spent months working with Huawei and Cambricon to rewrite parts of the model stack specifically for domestic Chinese hardware.
And here's a detail that raised eyebrows: back in February, Reuters reported that DeepSeek did not give Nvidia or AMD early access to V4 for optimization. Domestic suppliers got priority instead.
Counterpoint Research analyst Wei Sun told CNBC that V4's native support for Chinese chips could have "massive implications," helping Beijing achieve more AI sovereignty.
The market agreed immediately. Shares of Chinese chip manufacturers SMIC and Hua Hong Semiconductor surged 8.9% and 15.2% respectively in Hong Kong trading after the announcement.
For developers outside China, the practical question is straightforward: can you still run DeepSeek V4 on Nvidia GPUs? Almost certainly yes, given that it's open source and previous DeepSeek models ran fine on standard hardware. But the optimization priorities are clearly shifting toward domestic Chinese infrastructure first. That's worth keeping an eye on.
Why the Agent Optimization Matters More Than You Think
One detail in the release notes deserves way more attention than it's getting. DeepSeek said V4 has been specifically optimized for agent-based tasks and popular agent tools, including Anthropic's Claude Code and OpenClaw.
Neil Shah from Counterpoint Research called the DeepSeek V4 preview "a serious flex," pointing to lower inference costs paired with strong agent capabilities. Wei Sun added that V4's benchmark profile suggests "excellent agent capability at significantly lower cost."
Think about what that means in practice. This isn't just a bigger language model. DeepSeek is positioning V4 as a model purpose-built for the agentic coding workflows that are rapidly becoming standard in production environments. If you're building custom agents or using tools like Claude Code for autonomous coding tasks, V4 could be a real contender at a fraction of the cost.
For more on how agentic coding tools are evolving, check out our post on the current state of OpenClaw.
Benchmarks: Impressive, But Read the Fine Print
I want to be honest here. We're working with DeepSeek's self-reported numbers right now. Independent community benchmarks haven't been published yet, and that matters a lot.
That said, the claims are strong. V4-Pro beats all open-source models in math and coding. It's competitive with GPT-5.4 and Gemini 3.1-Pro. And according to the Stanford AI Index 2026, Chinese companies have "effectively closed" the AI performance gap with their American rivals overall.
But remember — this is a preview. Not a final release. DeepSeek is explicitly gathering real-world feedback before finalizing the model. Performance characteristics could shift before the stable version lands.
There's also a practical limitation worth noting. V4-Pro at 1.6 trillion parameters is prohibitively large for consumer hardware. You are not running this on a gaming rig. Period. The Flash variant at 284B is more accessible, but still demands serious infrastructure for local deployment. As the South China Morning Post noted, the extended technical report on V4's architecture is expected to benefit global AI developers, even if running the flagship model locally isn't feasible for most.
Should You Switch to DeepSeek V4 Right Now?
Short answer: no.
This is a preview release, not a production-ready model. The practical guidance from EvoLink's analysis earlier this month still holds — treat V4 as unreleased for production purposes until DeepSeek publishes official API model identifiers and finalized pricing.
What you should do right now:
- Download and evaluate. The model weights are open source. Pull them and test against your actual workloads.
- Build your eval set today. Use real tasks — bug fixes, refactoring, long-context code analysis — and benchmark V4 against whatever you're currently running.
- Don't bet on rumors. Wait for official API documentation, pricing pages, and stable model IDs before making production commitments.
- Watch for the technical report. DeepSeek's extended writeups on architecture and training have historically been gold for the research community, and V4's report should be no different.
The Bigger Picture: Open Source Keeps Closing the Gap
Zoom out for a moment. A year ago, DeepSeek R1 disrupted the entire conversation about whether you need billions of dollars to build frontier AI. Today, the DeepSeek V4 preview pushes that argument further.
An open-source model with 1.6 trillion parameters. Competitive with the best closed-source work from OpenAI and Google. Running on non-Nvidia hardware. At dramatically lower cost. The competitive dynamics just shifted again.
Bloomberg reported that V4 probably won't crash stock markets the way R1 did — investors have already priced in that Chinese AI is competitive and affordable. But Morningstar analyst Ivan Su made a sharp observation to CNBC: V4 directly positions itself against other Chinese open-source models. That framing didn't exist during R1's era, and it shows just how much domestic competition has intensified.
Meanwhile, several Chinese AI competitors took a hit in trading on Friday. MiniMax and Zhipu each fell around 8%, and Manycore Tech dropped 9%. DeepSeek isn't just challenging American labs anymore. It's reshaping its home turf too.
For a look at how these developments compare to what's happening on the other side of the Pacific, see our coverage of GPT-5.5 and what developers actually need to know.
Conclusion
The DeepSeek V4 preview release is a genuinely significant moment. The 1.6 trillion parameter V4-Pro and the cost-efficient V4-Flash — both open source, both with 1 million token context windows — represent a clear leap from V3. The Huawei chip optimization adds a geopolitical layer that developers and companies should be watching carefully.
But let's stay grounded. It's a preview. Test it, evaluate it against your real workloads, prepare your infrastructure, and wait for the production release before making switching decisions. The model looks extremely promising, and independent benchmarks from the community will tell us how much of that promise holds up.
If you're already experimenting with V4, I'd love to hear your early impressions in the comments. What tasks are you testing it on? How does it stack up against your current setup? Drop your findings below.
Sources
- CNBC — "China's DeepSeek releases preview of long-awaited V4 model as AI race intensifies" (April 24, 2026): https://www.cnbc.com/2026/04/24/deepseek-v4-llm-preview-open-source-ai-competition-china.html
- Al Jazeera — "China's DeepSeek unveils latest models a year after upending global tech" (April 24, 2026): https://www.aljazeera.com/economy/2026/4/24/chinas-deepseek-unveils-latest-model-a-year-after-upending-global-tech
- South China Morning Post — "DeepSeek unveils next-gen AI model as Huawei vows 'full support' with new chips" (April 24, 2026): https://www.scmp.com/tech/big-tech/article/3351239/deepseek-releases-next-gen-ai-model-world-leading-efficiency
- Reuters — "China's AI darling DeepSeek previews new model adapted for Huawei chip technology" (April 24, 2026): https://www.reuters.com/technology/chinas-deepseek-returns-with-new-model-year-after-viral-rise-2026-04-24/
- DW — "China's DeepSeek launches preview of new AI model" (April 24, 2026): https://www.dw.com/en/new-version-of-deepseek-ai-v4/a-76919223
- Bloomberg — "DeepSeek Unveils Flagship AI Model a Year After Breakthrough" (April 24, 2026): https://www.bloomberg.com/news/articles/2026-04-24/deepseek-unveils-newest-flagship-a-year-after-ai-breakthrough
- EvoLink — "DeepSeek V4 Release Date (April 2026 Update)": https://evolink.ai/blog/deepseek-v4-release-window-prep
- The Times of India — "China's DeepSeek launches V4 AI model" (April 24, 2026): https://timesofindia.indiatimes.com/technology/tech-news/chinas-deepseek-launches-v4-ai-model-claimed-to-outperform-google-gemini-chatgpt-and-other-american-ai-systems/articleshow/130483274.cms
- Stanford AI Index 2026 — Referenced via Al Jazeera reporting on US-China AI performance gap