On October 17, DeepSeek officially released the DeepSeek V4.1 Flash-1017 version, and the official web, App, and API models have been updated simultaneously. The core theme of this update is ‘Pareto improvement in performance-cost’: while model capabilities are comprehensively improved, both task runtime and output length are shortened, enabling users to complete tasks at lower cost and faster speed.
Benchmarks Leap Across the Board, Multiple Metrics Surpass Flagship Models
Compared with the previous V4.1 Flash, the 1017 version achieves significant improvements on multiple core benchmarks, with some metrics even surpassing flagship models with larger parameter counts.
Benchmark | V4.1-Flash-1017 | V4.1-Flash (previous version) | Kimi K3 | GPT 5.6-Sol | Claude Opus 5 |
|---|---|---|---|---|---|
HLE (no tools) | 42.6 | 36.8 | 43.5 | 44.5 | 56.3 |
Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 39.9 | 51.8 |
DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 73.0 | 74.0 |
ProgramBench | 29.1 | 20.3 | 17.5 | 23.0 | 37.0 |
NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 56.8 | 75.3 |
Automation-Bench | 60.5 | 54.8 | 46.7 | 45.8 | 50.3 |
Agents’ Last Exam | 35.4 | 31.8 | 27.6 | 26.7 | 28.6 |
Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 54.3 | 65.8 |
Terminal-Bench 4.0 was tested using the Claude Code framework, and DeepSWE using the Mini-SWE-V2 framework.
Most noteworthy is Terminal-Bench 4.0: it jumps from 31.2 to 42.3, not only greatly surpassing Kimi K3 (12.6) and GLM 5.3 (37.9), but even exceeding GPT-5.6-Sol’s 39.9. On the long-horizon coding benchmark DeepSWE v1.1, the 1017 version scores 77.0, overtaking Claude Opus 5’s 74.0, demonstrating strong competitiveness in agentic coding tasks.
‘Wishful Programming’: Building Complete Projects from Scratch
The core capability breakthrough in this update is that, through self-feedback-based reinforcement learning (RL) training, the model’s ability to build complete projects from scratch has been significantly improved, covering scenarios such as game development and 3D modeling.
Officials call it ‘Wishful Programming’: users only need to describe development requirements in natural language, and under conditions with extremely limited prompts and reference materials, the model can directly generate runnable project results. Official demo cases include 3D modeling of the Forbidden City, a web-based racing game, a Portal 2 recreation based on the Godot engine, and the 3D scene ‘Valley of Ages’ showing evolution from prehistoric times to the modern era.
Framework Adaptation: Special Optimization for Harness and Tencent Platforms
The new model has been specially optimized during training for platform frameworks such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy, improving task performance while shortening runtime and reducing token consumption. DeepSeek recommends that users upgrade Harness to version 0.2.1-rc.1 or higher for the optimal experience.
According to information released by the DeepSeek Harness team the same day, users logging in to use Harness for the first time will receive 6 yuan in bonus credit, with no recharge required. The bonus credit is issued in limited quantities and is valid for 7 days after arrival.
Search and Fetch APIs Open: 5 Yuan per Thousand Calls, Free for a Limited Time
Along with the new version release, the DeepSeek API officially opens the Search and Fetch interfaces, powered by DeepSeek’s self-developed search engine, which processes hundreds of millions of requests worldwide daily. Both interfaces are priced at 5 yuan per thousand calls.
It is worth noting that when users currently use the DeepSeek API in DeepSeek Harness, Claude Code, or Codex, they can already access these two interfaces through server-side tool calls, free for a limited time until November 1, 2026, after which per-call billing will begin.
Open Source Continuation: MIT License, Weights Also Released
The model weights of V4.1-Flash-1017 are now available for download on HuggingFace and ModelScope, continuing the V4.1 Flash strategy of open-sourcing under the MIT License. The API calling method and pricing remain unchanged; users can still access the latest model using `deepseek-flash`.
AIYXL Perspective:: The release of V4.1-Flash-1017 is a redefinition by DeepSeek of the positioning of its ‘Flash’ product line. In the past, ‘Flash’ meant ‘small, fast, and cheap, but with compromised capabilities’; the 1017 version proves that through RL training and framework-specific optimization, a small model can outperform flagship models on key tasks such as agentic coding while maintaining cost and speed advantages.
The phrase ‘Wishful Programming’ is worth pondering. It implies that the target users are no longer professional developers, but anyone who has ideas but cannot write code. When building a 3D scene or web game from scratch only requires ‘stating a wish,’ the threshold for AI programming is being redefined. DeepSeek chooses to open-source this capability under the MIT License and offer it at an unchanged price, continuing its consistent route of ‘technology democratization.’ At a time when GPT-5.6 and Claude Opus 5 occupy the performance high ground, DeepSeek is taking a differentiated path with a combination of ‘good enough + cheap enough + open enough.’