AI News Feed
Market watch
Trettitre Retro Ambient Audio System Ships After $2M KickstarterReport finds 135 'nudify' apps downloadable from 13-year-old Apple and Google accountsWindows 11 test build speeds up Edge launches, with wider app support plannedResearchers Study Whether Fighting Arctic Fires Can Protect Permafrost CarbonTim Cook visits Shanghai's Blue Box as brand runs local AI on Mac miniShaoshupai Relaunches FM3.14 Column With September Album Picks
Trettitre Retro Ambient Audio System Ships After $2M KickstarterReport finds 135 'nudify' apps downloadable from 13-year-old Apple and Google accountsWindows 11 test build speeds up Edge launches, with wider app support plannedResearchers Study Whether Fighting Arctic Fires Can Protect Permafrost CarbonTim Cook visits Shanghai's Blue Box as brand runs local AI on Mac miniShaoshupai Relaunches FM3.14 Column With September Album Picks
Large Language Models

DeepSeek V4.1 Flash 1017 builds 3D flying-Tesla demo for about 4 yuan, Ifanr reports

DeepSeek's V4.1 Flash 1017 built a 3D flying-Tesla demo for about 4 yuan in an Ifanr test and posted top coding results.

AI News Feed · 2026-10-19 19:11

Ifanr reported that V4.1 Flash's programming ability surpassed Kimi K3 and rose to first place among open-source models. On the DeepSWE benchmark, V4.1 Flash-1017 performed better than GPT-6.1 Sol, Claude Fable 5, Opus 5 and GPT-6 Astra, according to the report. In a coding-focused large-model arena, it reached what the report described as a new performance-cost Pareto optimum. On BuildingBench, a benchmark for 3D building reconstruction, its ranking rose from 13th in September to the top five, ahead of Gemini 4 Argon.

The flying-Tesla demo was run inside DeepSeek Harness from prompts alone, without uploading the original video. Ifanr said the project took 35 minutes and consumed 37.52 million tokens, including 330,000 output tokens, with a 99.6% cache hit rate. At peak-hour prices of 0.04 yuan per million cached input tokens and 8 yuan per million output tokens, the demo cost about 4 yuan. Ifanr estimated that achieving a similar result with GPT-6.1 Sol or Claude would cost at least 100 yuan, and that running the project on Codex with a ChatGPT Plus subscription could exhaust a five-hour quota without finishing while only using GPT-6.1 Max reasoning depth.

The report said the model also generated a 3D Eiffel Tower scene, a Zhangjiajie stone forest scene and an ocean-based space elevator inspired by The Wandering Earth. It replicated a hand-drawn first-person shooter built with Three.js. Using the REA reverse-engineering project, Ifanr had the model reconstruct a web version of the recently updated Typora 2.0 and fix a bug in its instant-rendering behavior. With HyperFrames, it also produced a 20-second doodle-style advertisement for the Mac mini; the report described the result as rough but capable of being improved with motion-design prompts.

DeepSeek said in its update notes that the new version uses self-feedback-based reinforcement learning to improve its ability to build games and 3D projects from scratch, according to the report. Ifanr noted that from September to October, several Chinese open-source model developers, including Kimi, MiniMax, Zhipu and Qwen, had not released new open-source models, while OpenAI, Claude, Gemini 4 and Meta continued to update or deploy models. The V4.1 Flash 1017 release shortens task runtime and output length, allowing users to complete tasks at lower cost and faster speed, the report said.