DeepSeek V4.1 Flash Hands-On: Open-Source Model Nears Closed-Source Flagships

Title

DeepSeek V4.1 Flash Hands-On: Open-Source Model Nears Closed-Source Flagships

Meta Description

DeepSeek^Ad V4.1 Flash 1017 tested: Terminal Bench score of 42.3% beats GPT-5.6 Sol, input at just $0.30 per million tokens, major gains in 3D generation and agentic coding, but speed and quality trade-offs remain.

Article Content

After a period of quiet, DeepSeek released V4.1 Flash (version 1017), positioning it officially as a "Pareto improvement"—better performance at equal or lower cost. According to a full hands-on test by Bilibili creator WorldofAI, the model performs strongly on terminal coding, 3D generation, and front-end development tasks, with some metrics even surpassing closed-source flagships like GPT-5.6 Sol. But it is not an entirely new architecture—it is an iterative upgrade of V4.1 Flash released on September 10, and DeepSeek has refined the series multiple times over the past two months.

Performance Leap: Benchmarks and Cost Control

The official benchmarks show considerable progress. On Terminal Bench 4.0, the score rose from 31.2% to 42.3%, surpassing GPT-5.6 Sol's 39.9%; on DeepSafe 1.1 it scored 77%, which according to DeepSeek community comparison data leads GPT-5.6 Sol, Claude Opus 5, and GLM 5.3. The HLE reasoning test improved from 36.8% to 42.6%, indicating a clear enhancement in reasoning capability.

On pricing, the model costs $0.30 per million input tokens and $1.20 per million output tokens, with a 1-million-token context window. WorldofAI commented that this makes it one of the best open-source models currently available, though the model is so large that it cannot realistically run locally. In WorldofAI's own benchmark, the model ranks around 10th to 11th overall—not in the top ten, but standout in agentic coding capability.

Terminal Bench 4.0 is an agentic coding benchmark set in a terminal (command-line) environment, testing whether a model can complete multi-step tasks in a real shell such as file operations, script writing, dependency installation, and debugging. Scores are typically measured by task pass rate. Unlike HumanEval-style benchmarks that focus on function-level code completion, it assesses long-horizon planning, tool calling, and error recovery, making it much closer to the real-world use case of "completing a project end to end."

HLE (Humanity's Last Exam) is a highly difficult interdisciplinary question set built jointly by academia and industry, covering expert-level problems in mathematics, physics, the humanities, and more. It was designed to provide discrimination after existing benchmarks become saturated. Its questions typically combine multi-step reasoning with obscure knowledge, and model scores are generally low—so an improvement from 36.8% to 42.6% is not small in absolute terms.

Agentic Capability: From Writing Code to Completing Projects

DeepSeek focused this release on agentic coding and project development, using reinforcement learning with self-feedback to improve the model's ability to complete full projects—whether generating games, building applications, or 3D modeling tasks. WorldofAI noted that this shift is significant: the model no longer just generates basic code but can complete complex development workflows end to end.

A supporting move is that DeepSeek opened its search and scraping APIs, backed by its own search infrastructure. The company says it can handle hundreds of millions of requests per day worldwide, with both endpoints billed at $0.75 per thousand calls. Search and scraping calls made through server-side tools such as CloudCode, Codex, or DeepSeek Harness are free until October 31, 16:00 UTC. The creator recommends using it with Harness, considering this the best combination to leverage the model's capabilities, because Harness optimizes tool calling, context management, and code execution around DeepSeek models, allowing the model to continuously test, identify errors, and iterate on output.

Reinforcement learning with self-feedback (self-feedback RL) refers to training where the model generates a complete solution and an execution environment or verifier provides automatically determinable feedback signals (such as whether tests pass, whether the program runs, or whether rendering succeeds), which is then used to optimize the policy without manual step-by-step labeling. Compared with sparse signals that only reward the final answer, this paradigm lets the model learn to self-trial, locate errors, and iteratively revise, and it is one of the mainstream approaches for improving long-horizon agentic task capability today.

Hands-On Performance: Blazing Fast, Strong in 3D and Front-End

During hours of intensive testing, WorldofAI found the model extremely fast, reported at around 244 tokens per second, with early tests even approaching 300 tokens per second. The creator said this may be faster than GPT-6.1 Sol's ultra-fast mode, while intelligence is on par with GPT-5.6 Sol, even surpassing it on some benchmarks.

3D generation is the area with the most obvious improvement. Compared with the original V4.1 Flash from two months ago, the quality of 3D scenes generated by the new version is described as "like night and day." In front-end development tests, layout, dynamic effects, subtle animations, and scroll triggers were all quite smooth. The output quality of a 360-degree product viewer was said to "imitate OpenAI's generation style," and the creator mentioned that the model sometimes identifies itself as an OpenAI model in certain deep-reasoning responses, prompting speculation about distillation training—he believes the DeepSeek team very likely distilled OpenAI models, but stressed that this is common practice among model providers.

In the classic pelican-riding-a-bicycle SVG test, with all models set to their highest reasoning level, Claude Opus was clearly best in visual style, while DeepSeek's output quality was acceptable. In another comparison with Opus 5.5, Opus won with better animation, sound effects, and detail, but DeepSeek cost only $0.23 versus Opus's $6—roughly a 26x cost difference; the trade-off was that DeepSeek took 1 hour while Opus needed only 20 minutes.

In a voxel art generation test, the model generated a medieval kingdom scene with over 2 million voxels in a single file, taking about 25 minutes at an API cost of just $5, and it even supported a day-night cycle and sidebar features such as adding flying dragons. In a comparison with GPT-6 Luna's 3D Tesla Roadster scene, DeepSeek cost $0.89, about 17% cheaper than Luna's $1.07, though the creator felt Luna's output quality was slightly better.

Model distillation refers to training a smaller model (the student) on outputs or distributions generated by a more capable large model (the teacher), so that the student approaches the teacher's performance at lower cost. Common industry practices include supervised fine-tuning on the teacher's generated reasoning traces or soft-label training on the teacher's output distribution. Because distillation relies on the teacher model's outputs, the student sometimes inherits the teacher's phrasing habits or even self-identity statements—which is exactly why the article's "model claims to be an OpenAI model" observation sparked speculation. It is worth noting that distillation itself is a widespread training method whose compliance is debated but which is widely used.

Strong Cost-Performance, But Not a Sweeping Lead

Cross-model comparisons reveal a more complex picture. In a single-HTML-file first-person parkour game test (4 stages, 1 round of fixes), DeepSeek's visuals were reported as the best, at a cost of $0.14, but it took 1 hour 30 minutes; Claude Haiku cost only 8 cents, took less time, and was actually better functionally. In a comparison with GPT-6.1 Sol Ultra, DeepSeek was about 90% cheaper ($0.44), the two took similar time, and DeepSeek could call web tools to gather reference material—a significant advantage when building complex projects.

The Minecraft sandbox comparison with Kimi K3 best illustrates the trade-off: Kimi took 2.5 hours and cost $35, delivering richer dimensions and graphical quality; DeepSeek took 1.5 hours and cost only $0.80—nearly 44 times cheaper—but struggled noticeably with many feature implementations and needed multiple rounds of prompting to get close. In an SVG test, Gemini 4 Argon finished in 13 minutes for $1.20, while DeepSeek took 50 minutes for only $0.40; the final images were close, with some water details even better, but Gemini's overall scene and speed were superior.

An even more extreme case was DeepSeek V4.1 with Hermes Agent calling Blender, taking about 7 hours to complete a 3D project at a cost of just $1.83. The creator suggests running such long tasks overnight. In video generation, the model also performed well, with the main bottleneck being the time required to generate assets.

Worth Trying, But Manage Expectations

WorldofAI's conclusion is that V4.1 Flash 1017 is one of the most impressive Flash-class models recently, with standout coding capability, 3D generation, and cost-performance, and it is currently available to try for free via the official chat. But it is not perfect: it can be slow at times, and quality on some tasks still lags top closed-source models, requiring multiple rounds of prompting to reach the desired result. For budget-sensitive developers and teams willing to accept longer wait times, this open-source model offers a highly attractive option—near-frontier output quality at a fraction of the cost of cutting-edge closed-source models.

One-line Summary

DeepSeek V4.1 Flash 1017 delivers major leaps in agentic coding and 3D generation at extremely low cost, but it does not lead across the board on speed and quality.

Paragraph Summary

DeepSeek released V4.1 Flash 1017, positioned as a Pareto improvement at the same cost. Official benchmarks show Terminal Bench 4.0 rising from 31.2% to 42.3%, beating GPT-5.6 Sol, and HLE reasoning rising from 36.8% to 42.6%; pricing is $0.30 input and $1.20 output per million tokens with a 1-million-token context. The model shifts focus to agentic coding, uses reinforcement learning with self-feedback, and opens search and scraping APIs ($0.75 per thousand calls). In testing it ran at about 244 tokens/sec, with notable gains in 3D generation and front-end development. Cross-model comparisons show its cost is often several to dozens of times lower, but it takes longer and some features require multiple prompting rounds—suitable for budget-sensitive teams that can accept the wait.