It's the first time in the history of the world.
DeepSeek is raiding this weekend!
Just recently, DeepSeek officials announced a major upgrade to V4.1 Flash, and DeepSeek-V4.1-Flash-1017 is fully up and running.
The core of this update is:It's a direct target of "Voice Programming!"
Currently, the DeepSeek web pages, apps, APIs are updated in sync, model prices remain unchanged, and weights are available for download.
In the benchmark test, the performance of the flagship version, V4.1-Flash-1017, took off in all directions.
In DeepSWE v1.1 programming, V4.1 Flash took the highest score of 77%, hitting GPT-5.6 Sol, Claude Opus 5.
What's even more terrifying is that it's scored well above the GPT-6 Astra (74%), while the average mission cost is only about $0.2.
In official terms, the upgrade achieved a "Pareto improvement" in performance and cost.
Furthermore, V4.1-Flash-1017 not only runs faster, works faster, and responds more precisely than the previous generation.
In DeepSeek's 3D modeling demonstration of the mansion, the footage crosses the red brick wall, overlooks the courtyard from a high elevation, and then turns to the city gate.
Buildings, squares, rivers and skies are composed of three-dimensional scenes that can be viewed from different angles.
The programming of vows, one key, comes true!
The new DeepSeek V4.1, the code is running.
After the Palace, DeepSeek released a web-based racing game.
At the start of the race, cars can be picked up, blue and purple racing cars and red racing cars can be picked up. At the end of the countdown, cars rush to the city streets illuminated by neon lights, with the speed chart on the right following changes.
Views can also be cut from the rear of the car into the cockpit. The steering wheel, rear-view mirrors, dashboards appear before the eyes, finish a race, and finally the settlement interface.
The selection of cars, races, drives, and settlements are combined into a process, demonstrating a race from start to finish.
This was followed by the Godot engine-based pump-transmission gate 2 torque.
In the white test box, the character moves in the first person, picks up the square, passes through a bright blue channel, and then moves to the next door.
The fourth project, Valley of Ages, shows the changes in cities from prehistoric to modern times.
In the Bronze Age, there were staircase buildings, in the Classical era there was a circular arena, in the Middle Ages there were walls and castles; later, factories, bridges, and high-rise buildings were connected.
The same scene, from prehistoric to modern. It presents historical change as a dynamic representation of the three dimensions.
DeepSeek attributes this improvement in development capabilities from zero toBased on feedbackStrengthening LearningTraining。
Reinforcement learning uses feedback to adjust model behaviour, with the official emphasis being on "self-feedback", hoping that the model will build a better complete project on tasks such as game development, 3D modeling and more.
In order for "voluntary programming" to be useful, the model has to make both scenarios, operations and feedback. After the individual developers have a working prototype, they can then change the gameplay, tweak the picture, and decide whether to continue with this point.
Programming attacks on GPT, Claude, upgrades without price
Among the new and old versions officially announced, Terminal-Bench 4.0 rose from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and DeepSWE v1.1 from 74.2 to 77.0.
In this official comparison table, the new versions of DeepSWE, Automation-Bench, Agents and Last Exam and Apex Bench scored higher than the other comparison models listed in the table.
Developers will also be interested in choosing a modelHow much does it cost to complete the same task?
DeepSeek also released a performance and cost chart for DeepSWE v1.1. The horizontal axis is the average cost per task, the longitudinal axis is the Pass@1 pass rate, and the testing framework is mini_swe_v2.
From the graph, the new version of the max file is approved at a rate of 77%, the average cost per task is between 1 and 2 yuan; the old version of the max file is approved at a rate of 74.2%, the average cost is between 3 and 4 yuan.
In other words, in this evaluation, the new version passed more tasks and the average cost per task was even lower.
DeepSeek refers to it as a "pareto upgrade for performance-enhancing costs" and says the new version improves the combination of performance and cost under multiple Harnesses.
The new model also shortened the runtime and output length of tasks, allowing users to complete tasks at a lower cost and faster speed, officials said.
The method of calling and the single price of the model API remained the same, and the new version can be called using deepseek-flash. The output is shorter, the task is completed faster, and the cost of tokens and waiting time are saved.
The actual cost of a programming agent also depends on how much output and how many attempts to make.It's hard to figure out how much the same task can be done, and how much money it will cost in the end.
Weightlifted open source, 100 billion web searches openDeepSeek has released the weighting addresses for both the Hugging Face and the ModelScope. The official version of the Hugging Face has been uploaded to 1017 and notes that both the code warehouse and the model weighting are licensed under the MIT license.
It's open and it's open.Self-developed search APIAccording to the official press release, the engine indexed 100 billion web pages globally, processed hundreds of millions of search requests per day, supported multiple languages, and filtered for SEO cheating, geo manipulation, and low-quality AIGC.
Open Search and Fetch are two interfaces, each responsible for finding the web page and crawling the web page, respectively. Developers can have Agent search and then read the original.
The two interfaces are priced separately.5 yuan / 1000 timesSuccessful and resulting searches are charged for valid seizures; failure, overtime or empty results are not charged.
There is another time point to remember.
Calls to service-end Search/Fetch tools initiated in DeepSeek Harness, Claude Code, Codex through the DeepSeek API, which are not yet billed, will be billed in installments starting at 00:00 PM Beijing time on November 1, 2026.
Once a prototype is made, it is often necessary to repeatedly change the gameplay, tweak the picture. Each change requires a model to be invoked, task costs are lower, and developers prefer to keep trying.
Weight is open, web pages, apps, APIs have been updated. You want to make a small game, build a 3D scene, and now you can test your points with the new version.
RoboInd editorial channelResearch, robotics companies and market intelligence for cross-border sourcing.Explore the channel ↗
Share this
End of editorialExplore more from Global Robotics News ↗