DeepSeek-V4-Flash Update

DeepSeek-V4-Flash Public Beta Release

DeepSeek has launched the public beta of the official DeepSeek-V4-Flash API. This update focuses on significantly enhancing agent capabilities, with benchmark results that exceed the previous V4-Pro-Preview. To use the updated model, developers simply need to set the model name to deepseek-v4-flash in their API calls.

Performance and Benchmark Results

DeepSeek-V4-Flash demonstrates substantial improvements in agentic and coding tasks. The following benchmark scores were reported:

  • Terminal Bench 2.1: 82.7
  • NL2Repo: 54.2
  • Cybergym: 76.7
  • DeepSWE: 54.4
  • Toolathlon verified: 70.3
  • Agent Last Exam: 25.2
  • Automation Bench (Public): 25.1
  • DSBench-FullStack (Internal): 68.7
  • DSBench-Hard (Internal): 59.6

For Code Agent tasks, the model was tested using the DeepSeek Harness minimal mode with max effort level, topp=0.95, and temperature=1.0.

Technical Specifications and Integration

DeepSeek-V4-Flash-0731 maintains the same model architecture and size as the DeepSeek-V4-Flash-Preview; the performance gains are the result of re-post-training.

Key integration details include:

  • Responses API: Native support for the Responses API format.
  • Codex Adaptation: Specifically adapted for Codex integration.
  • Scope: This update applies only to the deepseek-v4-flash API. The DeepSeek-V4-Pro API and APP/WEB models remain unchanged, with an official V4-Pro release expected soon.

Community Insights and Real-World Application

User discussions highlight the model's high intelligence-to-cost ratio, with many developers utilizing it as a primary "workhorse" for coding and agentic workflows.

Cost Efficiency and Performance

Developers report extremely low operational costs for high-volume tasks. One user noted spending only $4.55 USD for over 323 million tokens across 3,467 API requests in 30 days. Another user reported implementing complex tasks via MCP servers and context reducer tooling for approximately $0.50 per hour.

Use Case Specialization

Community feedback suggests a tiered model strategy where DeepSeek-V4-Flash is used for implementation and "grunt work," while more expensive models are reserved for high-level planning and reviewing:

"I've been driving flash model for 90% of my tasks... I try to keep changes under 1000 lines and drive architectural decisions myself, barely notice any difference compared to frontier models."

Local Deployment and Hardware

Because of its size and efficiency, the Flash model is seen as more accessible for local hosting compared to larger frontier models. Users have noted that it can run on prosumer hardware, such as dual RTX Pro 6000 GPUs, providing a fast and private alternative to cloud-based APIs.

DeepSeek API Evolution Timeline

DeepSeek has rapidly iterated its model offerings throughout 2024 and 2025:

  • July 2026: DeepSeek-V4-Flash public beta release with enhanced agent capabilities.
  • April 2026: Introduction of V4-Pro and V4-Flash, replacing legacy deepseek-chat and deepseek-reasoner names.
  • December 2025: Upgrade to DeepSeek-V3.2, introducing distinct thinking and non-thinking modes.
  • August 2025: DeepSeek-V3.1 release featuring a hybrid reasoning architecture and improved tool usage (SWE-bench Verified: 66.0).
  • January 2025: Launch of deepseek-reasoner (DeepSeek-R1).
  • December 2024: Upgrade of deepseek-chat to DeepSeek-V3.
  • August 2024: Implementation of Context Caching on Disk technology to reduce API pricing.

Sources