DeepSeek V4.1 Flash 1017 Hands-On: Open-Source Model's Coding and 3D Capabilities Approach Flagship Level

DeepSeek has once again stirred up the open-source AI community with a "minor version update." V4.1 Flash, first launched on September 10, has received the 1017 update, which the company positions as a "Pareto improvement"—higher performance at equal or lower cost. After a complete hands-on test, a Bilibili creator's verdict: this may be the strongest open-source model currently available, especially in three dimensions—agentic coding, 3D generation, and cost-effectiveness—where it far exceeds expectations.

A Non-Flagship "Pareto Improvement"

One fact needs to be clarified first: V4.1 Flash 1017 is not an entirely new flagship model, but a continuous iteration of the Flash series. DeepSeek^Ad says it tests better, responds faster, and outputs more concisely—essentially improving performance without raising costs. The model is already live on the DeepSeek platform, supporting both DeepThink and regular chat modes, and is simultaneously available via the mobile app and API.

The focus of this update is not general conversation, but agentic coding and project development. DeepSeek uses reinforcement learning with reflection to improve the model's ability to complete entire projects—shifting from merely generating basic code to handling complex development workflows end-to-end. Tasks like game development, app building, and 3D modeling are direct manifestations of this capability leap.

Benchmarks: Coding Ability Surpasses GPT 5.6

According to official and third-party data cited by the creator, V4.1 Flash 1017 shows clear improvements over the previous generation on multiple benchmarks:

• Terminal Bench 4.0 rose from 31.2% to 42.3%, surpassing GPT 5.6's 39.9%;
• On a certain coding-related benchmark, DeepSeek reports a score of 77%, exceeding GPT 5.6, Claude Opus 5.5, and GLM 5.3;
• HLE test scores improved from 36.8% to 42.6%, with logical reasoning capability advancing in tandem.

In the World of AI Benchmark's multi-domain comprehensive tests, the model currently ranks 10th or 11th. The creator's assessment: it has not yet entered the top ten, and gaps remain in some domains, but its coding ability is excellent. On pricing: $0.3 per million input tokens and $1.2 per million output tokens, with a context window of 1 million tokens. The model is too large to actually run locally, but as an open-source model called via API, this price is highly competitive.

Terminal Bench is a benchmark that evaluates whether AI models can complete tasks in real command-line environments. Tasks cover file operations, Git version control, system configuration, and script debugging, testing whether the model can translate natural language instructions into executable, verifiable terminal operation sequences. It is closer to real engineering scenarios than simple code completion, and is therefore often regarded as a hard indicator of agentic capability.

HLE stands for Humanity's Last Exam, a large-scale, high-difficulty academic question-answering benchmark launched in 2025. Questions span mathematics, physics, humanities, and engineering, authored by domain experts with anti-memorization-leakage mechanisms. It tests deep reasoning and knowledge integration capabilities. Scores are generally low, so every percentage point gained is significant.

Agentic Coding and Development Workflows

What really excited the creator was the model's performance in end-to-end development. Using a first-person parkour game as an example, four Flash models each built a game with four levels in a single HTML file. DeepSeek 4.1 Flash had the best visuals—although Hiku was more feature-complete and cheaper (8 cents vs 14 cents).

More noteworthy is the coordination with tool frameworks. The creator mentioned tools like DeepSeek Harness, purpose-built for this model, which optimize tool calls, manage context, execute code, and handle complex tasks. With the framework integrated, the model can continuously test, find errors, and improve results with tool assistance—rather than just generating code through a single API request. This combination further amplifies the model's agentic capabilities and is the key differentiator from "raw API calls."

Agentic coding refers to a model not just passively generating code snippets, but autonomously planning steps in a loop, calling tools (reading/writing files, executing commands, running tests), observing results, and correcting its approach accordingly until the task is complete. Compared with traditional "one-shot completion," it places much higher demands on context management, error recovery, and long-horizon planning, and is more likely to expose deviations across multiple iterations.

Harness in the AI engineering context is not the model itself, but an execution framework wrapping the model: it assembles prompts, schedules tool calls, maintains the context window, truncates or summarizes history, executes code in a sandbox, and collects error messages. The same model connected to different frameworks may perform noticeably differently, so when comparing model capabilities, it's worth checking whether framework conditions are consistent.

3D Generation: A Qualitative Leap from One Update

The creator compared 3D scene generation between the original V4.1 Flash released two months ago and the new model, concluding there is a "world of difference." A medieval kingdom scene created over 2 million voxels in a single file, completed in about 25 minutes at an API cost of just $5, and also implemented a day-night cycle and sidebar features like dragons.

Completing the task with Blender took about 7 hours at a cost of $1.83—the creator's take: suitable for running overnight, letting the model slowly complete tasks that require long polishing. Front-end development was equally impressive: typography, dynamic movement, micro-animations, and scroll-triggered details were "extremely smooth," and the 360-degree product viewer was stable and reliable.

The term "体速" appearing in the text should be voxel, i.e., a volumetric pixel in three-dimensional space, the basic building block of blocky 3D scenes like Minecraft. A scene containing millions of voxels means the model must organize massive amounts of geometric data and rendering logic in a single file while maintaining an interactive frame rate—a real test of code structure organization ability.

Blender is an open-source 3D creation suite covering modeling, materials, animation, and rendering. The article compares model generation with Blender production, essentially comparing the time and cost of two paths: "AI directly producing runnable code scenes" versus "manual modeling in professional tools." Note that the output forms differ, so the cost figures are of limited comparability.

Cost Comparison: A 26x Price Gap

Cost-effectiveness was a recurring theme in this hands-on test. With the same prompt, DeepSeek cost 23 cents while Claude Opus 5.5 required $6—about 26 times cheaper; DeepSeek took one hour while Opus took only 20 minutes, with comparable quality despite some flaws. In a comparison with GPT-6 Luna on a Tesla Roadster-themed 3D scene, DeepSeek cost about 89 cents vs Luna's about $1.70—roughly 17% cheaper—with Luna's quality slightly better.

The comparison with GPT-6.1 Soul Ultra was even more dramatic: DeepSeek cost about 90% less—just 44 cents—while achieving quite impressive results in roughly the same time. When co-creating a Minecraft sandbox with Kimi K3, Kimi took 2.5 hours and $35, while DeepSeek took only 1.5 hours and 80 cents, though there were functional issues. On generation speed, it's reportedly about 244 tokens per second, with early tests approaching 300 tokens.

Style Controversy and Accompanying Ecosystem

The creator noticed an interesting phenomenon: the model sometimes appears to claim it is an OpenAI AI model in certain responses, especially during deep logical reasoning. This sparked speculation about whether distillation was used in training, but he also emphasized this cannot be taken as conclusive evidence—"developers all do this, don't take this as absolute truth."

The accompanying tools are also worth noting. DeepSeek has opened its search and retrieval data API to everyone, based on its own search infrastructure, handling hundreds of millions of queries globally per day, with two endpoints at 75 cents per thousand calls. When used with Claude Code, Codex, or DeepSeek Harness, search and data retrieval through backend tools are currently free until October 31 at 16:00 UTC.

Distillation refers to training a smaller student model using the outputs (or probability distributions) of a more capable "teacher model"—a common knowledge transfer technique. A model claiming to be another company's product under specific prompts may be caused by that model's dialogue text being mixed into training data, or may stem from phrasing preferences during the alignment stage, and cannot alone serve as direct evidence of distillation. Judging such issues typically requires examining training data disclosures, tokenizer characteristics, or systematic comparisons of output distributions.

Worth Trying, But Don't Deify It

Overall, V4.1 Flash 1017 is indeed impressive in coding, 3D, and cost-effectiveness. In tasks like SVG generation, results are close to Gemini 4 Argon at 3x lower cost (40 cents vs $1.20), though it takes longer (50 minutes vs 13 minutes). It is still not perfect—quality and feature completeness on some tasks still lag behind top closed-source models. The creator's final advice is pragmatic: it's now free to try in chat and worth experiencing; if you're after ultimate quality, flagship models still have the edge; if you're after the balance of cost and results, this may be the strongest contender in the current open-source camp.

One-line Summary

DeepSeek V4.1 Flash 1017 achieves major leaps in coding and 3D generation at low cost.

Paragraph Summary

DeepSeek's V4.1 Flash, released on September 10, has received the 1017 update, which the company positions as a "Pareto improvement"—higher performance at equal or lower cost. According to a Bilibili creator's hands-on test, the model stands out in agentic coding, 3D generation, and cost-effectiveness: Terminal Bench 4.0 rose from 31.2% to 42.3%, surpassing GPT 5.6's 39.9%; HLE improved from 36.8% to 42.6%. 3D scene generation has qualitatively changed compared to the original version two months ago—a single medieval kingdom scene contains over 2 million voxels at a cost of just $5. The cost advantage is significant, about 26 times cheaper than Claude Opus 5.5 on similar tasks, but some tasks take longer, and quality and feature completeness still fall short of top closed-source models.