DeepSeek has released a preview of its new flagship model, called V4, and opened it for public use and download. The company announced the update on Friday last week and said both versions of the model are now live through its platform and API. This release follows months of quiet updates, after earlier hints appeared through new 鈥渆xpert鈥� and 鈥渇lash鈥� modes added to its system.
The model comes in two forms: V4 Pro is the larger system, built for coding and complex agent tasks. V4 Flash is smaller and faster (also cheaper) to run. DeepSeek said V4 Pro has 1.6 trillion total parameters with 49 billion active at any time, while V4 Flash runs on 284 billion total parameters with 13 billion active. The company said both models support long context and dual modes for reasoning.
This is the company鈥檚 biggest release since its R1 model in January last year. R1 gained global attention after showing impressive performance using limited computing resources. That earlier launch turned DeepSeek into one of China鈥檚 most talked about AI groups, according to MIT Technology Review.
V4 arrives during a time where things are pretty tense, with things like staff exits and delays to earlier launches. Even so, the release shows that DeepSeek continues work on advanced AI systems.
Why Is V4 Seen As So Powerful?
The strength of V4 comes from its performance with coding, maths and reasoning tasks. DeepSeek said V4 Pro beats all current open models in these areas and rivals leading closed systems. Benchmark results shared by the company show it performing at a level close to models such as Claude Opus 4.6, GPT 5.4 and Gemini 3.1.
MIT Technology Review reported that V4 Pro exceeds other open models like Qwen 3.5 and GLM 5.1 in coding and STEM tasks. It also performs well in writing and general knowledge tests, based on results released by DeepSeek.
The system also works well for multi step problems. DeepSeek said V4 Pro ranks among the top open models for agent based coding tasks. It has been tuned to work with tools such as Claude Code, OpenClaw, and CodeBuddy, which are used to build AI agents that can complete tasks on their own.
Developer feedback supports this and DeepSeek shared results from an internal survey of 85 developers, where more than 90% said they placed V4 Pro among their top choices for coding work. This shows how the model is already being used in real settings.
How Does It Deal With Huge Amounts Of Text?
One of the biggest features in V4 is its ability to process very long inputs. Both versions can handle up to 1 million tokens in a single prompt. MIT Technology Review said this is enough to fit the full text of The Lord of the Rings and The Hobbit together.
DeepSeek said this 1 million context length is now the default across its services. That puts it in line with the most advanced models currently available.
More from News
- Why Gen Z Is Choosing To Work At Startups Over Tech Giants?
- Instagram Just Launched A Disappearing Photo App In 2026 鈥� And Yes, That鈥檚 Exactly What It Sounds Like
- Inside VTEX Day 2026: Can The Brazilian Powerhouse Compete On The Global Digital Commerce Stage?
- Why Is The Education Sector Being Targeted In Cyber Attacks Lately?
- Experts Share: How Does The Tech World Feel About Apple鈥檚 New CEO
- What Do The April 2026 ONS Market Figures Mean For UK Businesses?
- FinanceWire And Symex Global Partner To Boost PR And IR Reach For Euronext Paris Companies
- Could You Be Answering A Normal Call When It鈥檚 Actually A Deepfake?
The real change is in how the model manages that volume of text. V4 uses a new attention system that compresses older information and keeps the most useful parts in focus. Instead of treating all past text equally, it selects what matters most and keeps nearby details in full.
This cuts computing costs, DeepSeek said V4 Pro uses 27% of the computing power needed by its earlier V3.2 model when handling a 1 million token prompt, and memory use drops to 10%. V4 Flash goes even lower, using only 10% of the computing power and 7% of the memory.
This makes it easier to build tools that need to read large datasets. That could include coding assistants that scan entire codebases or research tools that analyse long archives without losing track of earlier information.
Why Are Costs And Hardware Such A Big Deal Here?
V4 also is getting attention for its pricing. MIT Technology Review reported that DeepSeek charges $1.74 per million input tokens and $3.48 per million output tokens for V4 Pro. V4 Flash is priced at about $0.14 per million input tokens and $0.28 per million output tokens.
These prices are a lot lower than those of comparable models from OpenAI and Anthropic. This makes V4 one of the cheapest high level systems available, which could attract developers building apps and tools.
The model also shows a change in hardware use. V4 is the first DeepSeek model designed to work with Chinese chips such as Huawei鈥檚 Ascend series. Huawei confirmed that its Ascend 950 systems will support V4.
MIT Technology Review reported that this comes after US export rules limited China鈥檚 access to Nvidia鈥檚 top chips. DeepSeek has started using domestic chips for running the model, although parts of the training process may still use Nvidia hardware.
DeepSeek said costs could fall once Huawei鈥檚 systems are produced at scale later this year. If that happens, it could lower the price of running advanced AI and support the growth of a local AI tech stack in China.
