DeepSeek's V4 cuts running costs — but its Pro model still trails the best systems on independent tests. The company has previewed two open-source editions: V4 Pro with 1.6 trillion parameters and V4 Flash with 284 billion, both tuned for agent-style workflows and developer tools. Benchmarkers placed V4 Pro at 52 on an industry index (a rival scored 54), while analysts highlighted meaningful inference-cost improvements that could influence corporate AI procurement decisions.

Release, versions and openness

  • DeepSeek released a V4 preview in two editions: V4 Pro for heavier workloads and V4 Flash aimed at lower-cost use cases.
  • V4 Pro: 1.6 trillion parameters; V4 Flash: 284 billion parameters.
  • The company continues to publish open-source code, allowing enterprises and researchers to download, run and modify the models locally in most cases.
  • DeepSeek says V4 is tuned for compatibility with common agent frameworks and developer tools to ease integration into multi-step workflows.

Performance: clear progress, not a breakthrough

  • Independent benchmarks and hands-on reviews show V4 improves on DeepSeek's earlier V3.2 work but does not surpass the best open-source or closed-source offerings worldwide.
  • Artificial Analysis assigned V4 Pro a score of 52 on a 100-point-style index; a competing model scored 54 and several top closed-source models scored in the high 50s to 60s.
  • Strengths include better long-context processing, stronger agent-style inference and improved knowledge handling compared with the predecessor.
  • Persistent gaps remain in nuanced reasoning, creativity and certain code-generation tasks — areas important for developer copilot use and advanced automation.

Cost and agent advantage

  • Counterpoint Research highlighted V4's improved inference-cost profile versus earlier DeepSeek models, noting the cost gains are meaningful for agent-style workloads.
  • Lower inference costs can materially reduce monthly running bills for firms that deploy models at scale, affecting procurement calculations.
  • Because V4 is open-source, organisations that host the model on their own infrastructure can capture running-cost savings rather than paying per-request API fees.
  • Compatibility with common agent frameworks and developer tools helps firms integrate V4 into existing automation pipelines with less retooling.

Where V4 still falls short

  • Stress tests and third-party reviewers report uneven output quality on creative tasks and deficits on precision-sensitive benchmarks.
  • Benchmark comparisons place V4 below models such as Kimi K2.6 and several leading closed-source systems on code and reasoning metrics.
  • Even with a 1.6 trillion-parameter Pro edition, V4 remains behind top closed-source models on tasks requiring subtle judgement or high precision.

Related Articles

Artificial Analysis's 52 score and Counterpoint Research's praise for V4's cost profile point to modest progress: DeepSeek has narrowed running-cost disadvantages, but benchmarks show it still lags top closed-source and open-source rivals — a trade-off corporate buyers will have to weigh.

This article was created with AI assistance.