AI Weekly: Voice Input Tools Heat Up; SenseTime E-commerce Agent, DeepSeek V3.2 Released

Recently, AI products and models have seen a dense wave of updates, with new progress across voice input, e-commerce agents, open-source models, and AI hardware.

Voice Input Method Race Heats Up

ByteDance's Doubao Input Method has ended beta testing and officially landed in major app stores, available on both iOS and Android; WeChat Input Method continues to iterate on AI-assisted features; Zhipu, one of the 'Six Little Tigers' of large models, has also begun beta testing Zhipu AI Input Method.^[1]

Overseas, Wispr (Whisper Flow) recently announced the completion of a $25 million Series A round, bringing total funding to $81 million; Typeless topped Product Hunt on its launch day, and the team is also developing a mobile version.^[1]

In terms of product strategy, voice input tools can be roughly divided into desktop SaaS players, mobile big-tech players, and low-cost geek players. Doubao Input Method is considered standout in Chinese-language mobile scenarios, while Typeless has advantages in desktop output quality and transcription accuracy.^[1]

SenseTime Releases Ruying Marketing Agent

At its Product Launch Week, SenseTime released the Ruying Marketing Agent, moving into livestream e-commerce. It consists of 5 agents corresponding to five core roles: digital human host, live room control, livestream operations, store operations, and traffic advertising.^[2]

Among them, the livestream operations Agent can monitor data in real time and automatically generate multidimensional analysis reports, compressing review work that used to take about 2 hours into 20 minutes; the store operations Agent can automatically collect sales, traffic, and refund data, reducing a half-day process to 10 minutes. For complex 'refund analysis' tasks, it can automatically complete the download and integration of 7 spreadsheets, improving efficiency for a single task by 20x.^[2]

DeepSeek V3.2 Official Version Released

DeepSeek released the official version of V3.2 and V3.2-Speciale, focusing on improving reasoning and Agent capabilities. V3.2 is now available on web, App, and API, while the Speciale version is available as a temporary API.^[3]

The company says the official V3.2 integrates the thinking process into tool calling and supports tool calling in both thinking and non-thinking modes, reaching the highest level among current open-source models on agent tool-calling benchmarks. V3.2-Speciale's performance on reasoning benchmarks such as IMO 2025 and CMO 2025 is comparable to Gemini-3.0-Pro.^[3]

More AI Product Updates

Second Me is an AI identity built specifically for individuals, launched by Mindverse. By continuously learning a user's views, thoughts, memories, and ways of expression, it uses AI to fully preserve and extend a real 'me'.^[4]

After the global version of Looki L1 was released, it became popular in China, with average user usage time rising from 6.2 hours to 7.9 hours; Looki founder Sun Yang said the domestic version will be released soon.^[5]

In December, Founder Park recommended 7 new AI products, including flomo 5.0, Doka Camera, remio, Pallas AI, MuleRun, Odyss N1, and LavieAI, covering note-taking, cameras, personal AI partners, marketing, agent commercialization, food, and creativity.^[6]

Elena Vera, head of growth at Lovable, said Vibe Coding definitely has PMF, but Vibe Coding products have not actually found their own PMF, with core users changing every quarter; even though Lovable's ARR has exceeded $100 million, the definition of PMF changes every week.^[7]

In addition, recruitment at AI companies expanding overseas still faces challenges such as identifying overseas talent, employer brand operations, cross-border payroll and compliance, and remote team collaboration. Deel, which focuses on international recruitment, and AI recruitment platform Vorka.AI will host an online private workshop.^[8]