跪拜 Guibai
← All articles
Artificial Intelligence

Nvidia's NAS-Tuned Model Triples Throughput, Bragi Puts GPT-4.1 in Earbuds, and Alibaba Sunsets Virtual Numbers

By RTE开发者社区 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Nvidia's NAS-driven throughput gains and Qwen's sparse-activation approach both attack the same problem — inference cost — from opposite ends of the open-model spectrum. Bragi's white-label GPT-4.1 integration turns any headphone into an AI device without the brand needing its own model pipeline, which could accelerate voice-agent hardware faster than first-party efforts from Apple or Google.

Summary

Nvidia open-sourced Llama Nemotron Super v1.5, a 49B-parameter model derived from Meta's Llama-3.3-70B-Instruct that achieves SOTA on reasoning and agent benchmarks. Neural architecture search lets it run three times faster than its predecessor on a single GPU, cutting inference costs directly. Qwen3 also released a 30B model that activates only 3B core parameters in a non-thinking mode, targeting low-compute, high-performance deployments.

On the product side, Bragi's ChatAI platform packages GPT-4.1 for third-party earbud manufacturers. A multimodal local model preprocesses voice requests before they hit OpenAI's servers, and built-in speaker recognition suppresses false triggers. Bragi claims it stores no user data and is pursuing Apple MFi certification. Meanwhile, Microsoft's Edge browser gained an experimental Copilot Mode that replaces the new-tab page with a command input box and adds cross-tab agent actions, and Google's Search AI Mode now accepts image and PDF uploads alongside a Canvas planning sidebar.

Alibaba Communication announced it will stop renewals for its Ali Xiaohao virtual number service immediately and pull the app by October 31, 2025. Users who registered accounts with those numbers must migrate before the numbers are reclaimed.

Takeaways
Nvidia's Llama Nemotron Super v1.5 triples throughput over its predecessor and runs on a single GPU, using neural architecture search to balance accuracy and efficiency.
Qwen3-30B-A3B-Instruct-2507 activates only 3B of its 30B parameters in a non-thinking mode, targeting strong performance at very low compute cost.
Bragi's ChatAI platform lets any headphone brand integrate GPT-4.1 with on-device preprocessing, speaker recognition, and a no-server-logging privacy model.
Microsoft Edge's experimental Copilot Mode replaces the new-tab page with a command input and adds cross-tab agent actions plus voice navigation.
Google Search AI Mode now supports image and PDF uploads for queries, a Canvas planning sidebar, and real-time visual conversations via Lens.
Eigent, the first open-source desktop multi-agent system from CAMEL-AI, runs customizable agent workforces locally on MacOS and Windows with over 200 built-in MCP tools.
China Telecom's TeleAI GLASSES pack a 12MP camera, five-mic array, and the Xingchen large model into 38-gram frames priced at 1,999 RMB.
Alibaba's Ali Xiaohao virtual number service stopped renewals on July 29, 2025, and its app will be removed October 31; users must unbind linked accounts before numbers are reclaimed.
WAIC 2025 in Shanghai drew over 800 exhibitors and 1,500 experts, with 31 signed projects totaling more than 15 billion RMB in investment.
Former Google CEO Eric Schmidt stated at WAIC that open-source models are a defining feature of current AI development and called for cross-border researcher collaboration on value alignment.
Conclusions

Bragi's architecture — local multimodal preprocessing before any data hits OpenAI — is a practical privacy pattern that sidesteps the 'always-listening' fear that has stalled always-on voice assistants.

Nvidia's NAS-based throughput improvement and Qwen's sparse activation are two different answers to the same question: how to make large models cheap enough to run locally. The former optimizes the full model; the latter only lights up the parts it needs.

Microsoft's Copilot Mode and Google's AI Mode are converging the browser into an agent runtime. The new-tab page disappearing into a command box signals that the URL bar is no longer the primary interface.

Alibaba killing Ali Xiaohao with only three months' notice creates a real operational headache for anyone who used those virtual numbers as two-factor authentication or account recovery phone lines.

Concepts & terms
Neural Architecture Search (NAS)
An automated method that searches for the optimal neural network structure — layer types, connections, widths — to maximize a given objective like accuracy or speed, rather than relying on hand-designed architectures.
Non-thinking mode
Qwen's architecture innovation where only a small subset of a model's total parameters (3B out of 30B) are activated during inference, trading off some reasoning depth for dramatically lower compute cost and higher speed.
MCP (Model Context Protocol)
An open protocol that standardizes how AI agents connect to external tools and data sources, analogous to a USB-C for agent-tool integration. Eigent's system ships with over 200 pre-built MCP tools.
Full-duplex audio
The ability to send and receive audio simultaneously, like a phone call, rather than walkie-talkie-style half-duplex where only one side can speak at a time. TeleAI GLASSES supports this for natural conversation flow.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗