Today,DeepSeek-V4has finally arrived.
The preview is now live and open source.
DeepSeek released two editions:
DeepSeek-V4-Pro and DeepSeek-V4-Flash.
Both support a one-million-token context window, with simultaneous updates across the website, app, and API.
Pro targets frontier performance; Flash is smaller, faster, and more cost-efficient.
Both editions support standard and reasoning modes.
Reasoning mode also lets users set reasoning_effort; max is the recommended setting for complex agent workflows.
Existing deepseek-chat and deepseek-reasoner users have a three-month migration window.
V4-Pro: Performance on Par with Leading Proprietary Models
DeepSeek highlights three major gains:
1. A substantial leap in agent capability:
V4-Pro moves well beyond its predecessor.
The improvement is especially clear in agentic coding, where it reaches the top tier of open-source models and performs strongly across several other agent benchmarks.
Internal adoption offers another practical signal: DeepSeek employees are already using it as their everyday agentic coding model.
They report a better experience than Sonnet 4.5 and delivery quality broadly comparable to Opus 4.6 without reasoning, although Opus 4.6 in reasoning mode still holds an advantage.
2. Broad world knowledge:
DeepSeek-V4-Pro significantly outperforms other open-source models on world-knowledge evaluations, trailing only the leading proprietary Gemini 3.1 Pro.
3. Frontier reasoning performance:
Across mathematics, STEM, and competitive programming benchmarks,
DeepSeek-V4-Pro outperforms every publicly evaluated open-source model
and delivers results comparable to the world's strongest proprietary systems.
Those claims are impressive, but they still leave the most important question unanswered: what becomes possible in practice?
The clearest place to start is the one-million-token context window.
This matters more than it may appear.
A year ago,a one-million-token context window was a premium feature available in only a handful of models.
DeepSeek has now made it standard across allofficial servicesand released it as open source.
This is not simply another model claiming a million-token window. DeepSeek is moving long context toward broader access and infrastructure-level availability.
The technical direction is equally clear. V4 introduces a new attention mechanism that compresses at the token level, then combines it with DSA sparse attention to reduce the compute and memory cost of long-context workloads.
That is more than an incremental upgrade.
Agents, codebase analysis, long-document processing, and complex workflows all face the same constraint:insufficient context.
Real business work brings documents, logs, message history, tool outputs, and entire file trees into the same operating context.
Without enough context capacity, many valuable workflows cannot be built reliably.
By making one-million-token context a baseline rather than a showcase feature,
V4 delivers more practical value than a marginal benchmark gain.
It turns a capability previously reserved for a few premium models into something teams can realistically build on.
Next comes agent performance.
V4 has been adapted and optimized for widely used agent products including Claude Code, OpenClaw, OpenCode, and CodeBuddy, with improvements across coding and document-generation tasks.
Scroll to explore
Sample presentation slide generated by V4-Pro in an agent framework
Pricing
The DeepSeek API now offers both V4-Pro and V4-Flash.
It supports the OpenAI Chat Completions API and the Anthropic API.
To access the new models,
the base_url remains unchanged; set the model parameter to deepseek-v4-pro or deepseek-v4-flash.
V4-Pro currently costs RMB 12 per million input tokens and RMB 24 per million output tokens.
V4-Flash costs RMB 1 for input and RMB 2 for output.
By DeepSeek's own historical standards, this is not the kind of disruptive pricing that instantly resets the market.
Against today's competitive landscape, however—especially compared with leading international models—it remains highly attractive.
More importantly, Flash is not a token entry-level edition.
It can complete real tasks, support straightforward agent workflows, and still provide a one-million-token context window at a markedly lower cost.
Domestic compute support is another strategic part of the release.
DeepSeek's announcement does not disclose every implementation detail, but the external signals are clear.
Reuters reports that the new model has been adapted for Huawei Ascend technology.
Huawei has also said its Ascend 950-based supernode will fully support V4.
Reuters interprets this as a meaningful shift away from DeepSeek's previous dependence on NVIDIA and toward China's domestic AI chip ecosystem.
Why does that matter?
Large-model competition is no longer only about models. Compute, chips, deployment architecture, and supply chains are all part of the equation.
A powerful model locked into a constrained hardware ecosystem faces limits that extend far beyond technical performance.
There is, however, one clear omission.
V4 does not yet include multimodal capability.
Multimodality has become close to a baseline expectation.
Without vision, a model cannot interpret images, leaving a structural gap in many agent scenarios.
Screenshot analysis, interface understanding, computer use, and visual reasoning are all affected.
V4 remains text-only, and that will disappoint some users.
Seen from another angle, the decision also reveals a disciplined set of priorities.
DeepSeek is strengthening the foundation first:
long context, agents, open source, hardware compatibility, production APIs, and viable economics.
Multimodality has not been played in this release.That does not mean it is absent from the roadmap; it may simply not be the capability DeepSeek is ready to lead with today.
V4 may not feel revolutionary at first glance,
but it represents a meaningful step toward a more complete and commercially realistic path for China's large-model ecosystem.
DeepSeek closed its announcement with a line from the philosopher Xunzi:
Do not be seduced by praise or intimidated by criticism. Follow the right path and hold yourself upright.
The principle is relevant well beyond this release.
Technology creates lasting value not by appearing more powerful,
but by integrating into real operations, systems, and data until it becomes a dependable business capability.
That is the work GeekOnUp continues to pursue.
We follow every major shift in the model landscape, but we are less interested in short-lived hype than in durable outcomes.
Our focus is clear:which capabilities can solve real enterprise problems today, and which foundations will continue to compound through the next wave of change.
In an industry moving this quickly,the best response is to execute rigorously and build on solid ground.
Progress does not need to be rushed. What matters is building the depth to keep moving forward.
