DeepSeek and MiniMax Update Flash Models

Chinese AI company DeepSeek released an updated version of its existing model, DeepSeek-V4.1-Flash-1017, on October 17, 2026. MiniMax announced MiniMax M3.1 Flash on October 19. Both companies highlighted improved performance in AI agents and coding, along with reductions in processing time or usage costs.

DeepSeek said the update improved results on evaluations for reasoning, coding, agent tasks, and visual understanding, while reducing the time and output needed to complete tasks. It reported that its software development score on DeepSWE v1.1 rose from 74.2 to 77.0, and its Terminal-Bench 4.0 score increased from 31.2 to 42.3. The company also said reinforcement learning using self-feedback improved the model’s ability to build projects from scratch.

DeepSeek gave examples including a 3D model of the Forbidden City, a browser-based racing game, and a game built with the Godot engine. The model is available through the web version, app, and API, and has also been optimized for DeepSeek Harness and Tencent’s CodeBuddy and WorkBuddy. Its model weights are available on Hugging Face and ModelScope under the MIT license.

MiniMax M3.1 Flash is designed to improve performance in agent tasks, coding, PC operation, and cybersecurity. MiniMax said the anonymously available “Space Bunny” was an early version of the model. It reported that online inference speed is about 3 times faster than before and that total usage costs for agent tasks have fallen 70%. For everyday use, it said output speed exceeds 200 tokens per second, with a median above 170 tokens per second.

MiniMax identified enterprise workflows across multiple SaaS products, multistep planning, and maintaining consistency across documents and files as intended use cases. It reported scores of 61.7 on OSWorld V2 for PC operation and 82% on CyberGym for security. The model supports a context length of 1 million tokens and accepts text, image, and video inputs. Its reasoning depth can be set across 5 levels to limit processing time or output token counts for different uses.

Pricing per 1 million tokens is $0.15 for input, $0.60 for output, and $0.015 for cache reads. MiniMax began offering the model through the MiniMax API, its M Plan subscription, and MiniMax Code; the company plans to release the model weights later.

Chinese AI company DeepSeek released an updated version of its existing model, DeepSeek-V4.1-Flash-1017, on October 17, 2026. MiniMax announced MiniMax M3.1 Flash on October 19. Both companies highlighted improved performance in AI agents and coding, along with reductions in processing time or usage costs.

DeepSeek said the update improved results on evaluations for reasoning, coding, agent tasks, and visual understanding, while reducing the time and output needed to complete tasks. It reported that its software development score on DeepSWE v1.1 rose from 74.2 to 77.0, and its Terminal-Bench 4.0 score increased from 31.2 to 42.3. The company also said reinforcement learning using self-feedback improved the model’s ability to build projects from scratch.

DeepSeek gave examples including a 3D model of the Forbidden City, a browser-based racing game, and a game built with the Godot engine. The model is available through the web version, app, and API, and has also been optimized for DeepSeek Harness and Tencent’s CodeBuddy and WorkBuddy. Its model weights are available on Hugging Face and ModelScope under the MIT license.

MiniMax M3.1 Flash is designed to improve performance in agent tasks, coding, PC operation, and cybersecurity. MiniMax said the anonymously available “Space Bunny” was an early version of the model. It reported that online inference speed is about 3 times faster than before and that total usage costs for agent tasks have fallen 70%. For everyday use, it said output speed exceeds 200 tokens per second, with a median above 170 tokens per second.

MiniMax identified enterprise workflows across multiple SaaS products, multistep planning, and maintaining consistency across documents and files as intended use cases. It reported scores of 61.7 on OSWorld V2 for PC operation and 82% on CyberGym for security. The model supports a context length of 1 million tokens and accepts text, image, and video inputs. Its reasoning depth can be set across 5 levels to limit processing time or output token counts for different uses.

Pricing per 1 million tokens is $0.15 for input, $0.60 for output, and $0.015 for cache reads. MiniMax began offering the model through the MiniMax API, its M Plan subscription, and MiniMax Code; the company plans to release the model weights later.