The official V4 PRO version is seven months in the making—can DeepSeek make a splash in the programming agent arena?

According to ifeng.com, on July 31, the official API for DeepSeek V4 Flash was launched for public beta testing.

The released model, versioned V4-Flash-0731, features a total of 284 billion parameters with 13 billion activated, utilizing a Mixture of Experts (MoE) architecture and supporting a 1 million token context window.

Compared to the preview version released in April, the model structure and parameter scale remain identical, with only post-training adjustments made.

According to official benchmark data, the Flash official version outperformed the V4-Pro preview version released in April across nine Agent evaluations.

Notably, the DeepSWE score for real-world software engineering tasks rose from 7.3 in the Flash preview version to 54.4 in the official version.

In terminal operation testing (Terminal Bench 2.1), the score reached 82.7, second only to Opus 4.8's 85.0 among publicly reported models, and higher than GLM-5.2's 81.0 and the V4-Pro preview version's 72.1.

Cybersecurity attack-defense simulation (Cybergym) scores improved from 38.7 to 76.7. The comprehensive tool invocation test (Toolathlon-Verified) rose from 49.7 to 70.3.

In the Artificial Analysis intelligence index evaluation, the Flash official version scored 50 at the highest reasoning tier, compared to a median of 17 among comparable models.

Also unveiled is the minimalist mode of the AI Agent evaluation framework, DeepSeek Harness. The official changelog notes that all Code Agent tasks in the public benchmarks were run using this framework, set to the max parameter tier.

Reportedly, this upgrade only involves the Flash API interface. The web, App, and V4-Pro API interfaces remain unchanged. Regarding the official V4 Pro version, the company stated it will be released "as soon as possible."

01 DeepSeek V4's Development Spans Nearly Seven Months

Reviewing the release timeline for DeepSeek V4, official version announcements were made in January, March, June, and July of this year, though none materialized as scheduled. The most recent announcement came on July 28, when media revealed that DeepSeek's self-developed programming tool, Harness, was preparing for closed beta testing, with initial testers to be selected from questionnaire respondents.

Reports also disclosed that the V4 official version would launch one to two weeks after Harness's internal testing, suggesting a window between August 10 and 20. If accurate, the V4 version's development timeline from initial announcement to release would exceed seven months.

The performance of the V4 Flash official version has objectively raised market expectations for the Pro official version. In comparison, Flash has 284 billion total parameters (284B) with 13 billion activated (13B), while Pro has 1.6 trillion total parameters (1.6T) with 49 billion activated (49B), marking a significant gap in parameter scale.

Moreover, in performance, during the Terminal Bench 2.1 benchmark—which evaluates an AI Agent's ability to autonomously plan, execute multi-step command-line tasks, and recover from errors in a terminal environment—the V4 Flash official version scored 82.7, compared to the Pro preview version's 72.1. Given that the Pro official version is positioned and priced higher than Flash, the extent of quantifiable improvements over Flash's performance has become a market focus.

02 What Is DeepSeek Waiting For?

Among the data released for the Flash official version, Agent capabilities represent the core focus of the upgrade. According to official benchmarks, Flash scored 82.7 in Terminal Bench 2.1, with DeepSWE rising from 7.3 to 54.4, surpassing the Pro preview version across all nine Agent evaluations. The testing framework used was DeepSeek Harness in minimalist mode, marking Harness's first official appearance as a product since DeepSeek's bet on code agents.

Public reports indicate that following the global success of Anthropic's Claude Code desktop agent programming product, DeepSeek internally formed the Harness code agent team in March 2026, adhering to the principle of "Model + Harness = Agent," where Harness is defined as the "model control layer" connecting large models with agent products.

After securing 50 billion yuan in initial external funding in June 2026, Harness was explicitly listed as a core investment area, prompting large-scale recruitment and repeated reports of persistent staffing shortages. On July 23, Liang Wenfeng clarified priorities in a closed-door investor meeting, emphasizing a focus on the Coding Agent.

From a product perspective, Harness aims to deliver a complete engineering agent capable of entering terminals, understanding project structures, reading and writing files, executing commands, autonomously fixing errors, and managing version control. This represents a clear strategic extension from DeepSeek's previous model weight open-source focus, seen as a shift from providing "a brain" to providing "a brain plus hands."

Notably, during the window when DeepSeek committed resources to code agents, the industry validated similar judgments at a faster pace.

In overseas markets, OpenAI completed a full turnaround in the programming arena within six months. In September 2025, its programming tool, Codex, had only about 5% of Claude Code's usage. Company president Greg Brockman later admitted to "learning lessons too late."

Subsequently, OpenAI embedded Codex into ChatGPT's desktop and mobile versions, leveraging its user base of hundreds of millions for rapid expansion, and introduced one-click migration from Claude Code. By July 2026, Codex's weekly active users exceeded eight million. On the same day as the Flash official version's launch, OpenAI announced an 80% price cut for GPT-5.6 Luna.

The Chinese market saw equally fierce competition during this period. Alibaba's Qoder ranked first with a 47.6% share in IDC's 2025 China AI Programming Tools Market Report. In February, Alibaba Cloud Bailian reduced Coding Plan pricing to 7.9 RMB per month, followed by nine platforms offering 28 different pricing tiers; Zhipu's GLM Coding Plan attracted over 240,000 global paid developers; ByteDance's TRAE expanded from an IDE to a full-platform tool and open-sourced its Agent toolkit; Kimi K3 topped the Arena AI Code rankings.

Capital market investments in the sector have been equally swift. In June, SpaceX acquired Anysphere, the parent company of the AI-native code editor Cursor, in a $60 billion all-stock deal. Cognition, the parent company of the world's first "fully autonomous AI software engineer," Devin, secured over $1 billion in funding the same month, valuing the company at approximately $26 billion. The programming tool Windsurf was split through competitive bidding among Google, Cognition, and OpenAI—Google acquired the core technology and founding team for about $2.4 billion, while Cognition took over the product and clients. Multiple leading companies bidding for a single programming tool in the same window indicates that capital now prices programming agents as a standalone sector.

For DeepSeek, the competitive pressure in the track (translated as "arena" or "sector") is undeniable. Over the past year, the programming agent market has continuously raised entry barriers, meaning the later Harness launches, the greater the product, channel, and user habit gaps it must overcome.

Thus, even though the V4 Flash official version has proven the competitiveness of its model foundation and post-training methods, DeepSeek remains highly cautious about releasing Harness. The confidentiality agreement in Harness's internal testing, which states that "leaks will affect future eligibility for DeepSeek's various model, product internal tests, and cooperation opportunities," reflects both technological protection and product rhythm control.

Nevertheless, with the release of the V4 Flash official version, the moment of truth for DeepSeek's ability to make a splash in the programming agent arena is drawing ever closer.

(Image credits and public data sources for this article are from the internet and public industry platforms. If there are any infringements, please contact us promptly for corrections.)