📜 DeepSeek API Evolution & Change Log

🚀 Latest Update: DeepSeek-V4-Flash

Date: July 31, 2026

The public beta for the deepseek-v4-flash API is officially live. To integrate this version, users simply need to update their model parameter to deepseek-v4-flash; no other changes to the calling method are required.

🤖 Agentic Performance

This release introduces massive leaps in agent capabilities, significantly outperforming the V4-Pro-Preview.

Technical Notes on Testing:

1. For public Code Agent benchmarks, the model utilized the DeepSeek Harness minimal mode (upcoming release) with the following hyperparameters:
   • Effort: $\text{max}$
   • Top-p: $topp = 0.95$
   • Temperature: $T = 1.0$
2. DSBench-FullStack and DSBench-Hard are proprietary internal evaluations for full-stack and complex coding tasks, respectively.

🛠️ Integration & Architecture

• Compatibility: Native support for the Responses API and specialized optimization for Codex.
• Configuration: Detailed settings are available in the [official documentation].
• Build: DeepSeek-V4-Flash-0731 retains the same size and architecture as the Preview version, having undergone only re-post-training.
• Stability: No changes were made to the DeepSeek-V4-Pro API or the Web/App interfaces.

---

📅 Historical Versioning

DeepSeek-V4 Transition

Date: April 24, 2026

The API now supports both deepseek-v4-pro and deepseek-v4-flash. These are accessible via Anthropic or OpenAI ChatCompletions interfaces using the standard base_url.

deepseek-chat and deepseek-reasoner are now legacy and were scheduled for discontinuation on 2026-07-24. During the transition:

• deepseek-chat $\rightarrow$ deepseek-v4-flash (Non-thinking)
• deepseek-reasoner $\rightarrow$ deepseek-v4-flash (Thinking)

The V3.2 Era

Date: December 1, 2025

• deepseek-chat $\rightarrow$ V3.2 (Non-thinking)
• deepseek-reasoner $\rightarrow$ V3.2 (Thinking)

🌟 DeepSeek-V3.2-Speciale

A limited-time version was deployed via a specific endpoint: base_url= https://api.deepseek.com/v3.2_speciale_expires_on_20251215

• Price: Same as V3.2.
• Limitation: No tool calling.
• Expiry: Dec 15, 2025, 15:59 UTC.

Date: September 29, 2025

• Both chat and reasoner endpoints were upgraded to DeepSeek-V3.2-Exp.

---

V3.1 Iterations

Date: September 22, 2025 (Terminus Update) Upgraded to DeepSeek-V3.1-Terminus. This update focused on refinement:

• Language Consistency: Fixed abnormal characters and reduced "Chinglish" (CN-EN mixing).
• Agent Optimization: Boosted Search and Code Agent performance.

Date: August 21, 2025 (V3.1 Launch) Introduction of a Hybrid Reasoning Architecture, allowing one model to toggle between thinking and non-thinking modes.

• Efficiency: Faster response times than DeepSeek-R1-0528.
• Agent Benchmarks:
  • SWE-bench Verified: $66.0$
  • SWE-bench Multilingual: $54.5$
  • Terminal-bench: $31.3$

---

Early 2025 Enhancements

deepseek-reasoner (May 28, 2025)

Upgraded to DeepSeek-R1-0528.

• Reasoning Gains ($\Delta$):
  • AIME 2025: $70.0 \rightarrow 87.5$ ($+17.5$)
  • GPQA: $71.5 \rightarrow 81.0$ ($+9.5$)
  • LCB_v6: $63.5 \rightarrow 73.3$ ($+9.8$)
  • Aider: $57.0 \rightarrow 71.6$ ($+14.6$)
• Other Improvements:
  • UI/UX: Better aesthetics for generated games/web pages.
  • Reliability: Drastic reduction in hallucinations.
  • Tooling: JSON output and Function Calling (Tau-bench: 53.5 Airline / 63.9 Retail).

deepseek-chat (March 24, 2025)

Upgraded to DeepSeek-V3-0324.

• Performance Boosts:
  • MMLU-Pro: $75.9 \rightarrow 81.2$
  • GPQA: $59.1 \rightarrow 68.4$
  • AIME: $39.6 \rightarrow 59.4$
  • LiveCodeBench: $39.2 \rightarrow 49.2$
• Key Features:
  • Writing: Enhanced Chinese proficiency, aligning with R1 style for long-form content.
  • Search: More comprehensive report analysis.
  • Dev: Higher accuracy in front-end code generation.
  • API: Fixed Function Calling bugs from previous V3 iterations.

---

Foundation Releases

Date: January 20, 2025

• deepseek-reasoner launched, powered by DeepSeek-R1.

Date: December 26, 2024

• deepseek-chat upgraded to DeepSeek-V3.