lunin

Qwen3-ASR can decipher speech, songs, and even rap, coping with low-quality recordings and noise.

11 languages ​​are supported: English, Chinese, Arabic, Spanish, Korean, Russian, etc. You can specify a list of names and keywords so that they are recognized correctly.

Cost - $0.000032 per second of audio via API.

Demo: https://huggingface.co/spaces/Qwen/Qwen3-ASR-Demo

Alibaba Releases Qwen3-ASR - Speech Recognition Model

The company unveiled Rubin CPX, a graphics processor optimized for tasks with a context of over 1 million tokens.

The chip is designed for “disaggregated inference,” an approach where different GPUs process different parts of a task. This should improve the efficiency of models in video generation, programming, and other long-context scenarios.

Rubin CPX will be released in late 2026.

Nvidia Shows Off GPU for Ultra-Long Context Models

Anthropic has released a significant update to Claude, enabling the chatbot to generate and edit spreadsheets, documents, PowerPoint slides and PDF files. Users can upload raw data and describe their desired outcome, leaving Claude to handle the formatting and file creation.

This feature is currently available for Max, Team and Enterprise plans, and will be made available to Pro users in the coming weeks.

More details: https://www.anthropic.com/news/create-files

Claude can now create spreadsheets and presentations

Although Qwen3-Max-Preview cannot reason, it produces good results in programming and scientific disciplines, particularly mathematics. In tests, the model outperformed both DeepSeek V3.1 and Claude Opus 4 (in non-reasoning mode).

Qwen3-Max-Preview is also the first Qwen model with a trillion parameters.

Try it here: https://chat.qwen.ai/

The flagship model, Qwen3-Max-Preview, has been released

The tool analyzes long videos, finds the moments you need and assembles a video with subtitles from them.

![](https://m.stacker.news/106824)

Genspark has released an AI editor for video editing

nostr

Decentralized Social Network Platform Goes Live with Full Support and Backwards Compatibility

Nostria, the innovative Nostr protocol client designed for global scale, today announced the official launch of its Minimum Viable Product (MVP), marking a pivotal milestone in decentralized social networking. As planned, the MVP was completed on August 31st, 2025, transitioning from development phase to a fully supported platform with guaranteed backwards compatibility going forward.

https://www.nostria.app/press/mvp-released

Nostria launch: The Beginning has Started

Since September 1, China has made it mandatory to label all content created with AI - from text and images to audio and video. Now, any social network is required to put visible labels like “AI-generated,” as well as hidden ones - in the form of watermarks or metadata.

Platforms like WeChat, Douyin, Weibo, and RedNote have already implemented new rules: authors are forced to indicate the use of AI, and attempts to remove the label can result in the account being blocked. Weibo even allows users to complain about posts that do not indicate that they were made by a neural network.

China’s social media sites rush to abide by AI-generated content labelling law

Relatively unknown startup Nous Research has released Hermes 4, a new family of open-source models that rivals leading competitors in terms of performance while avoiding heavy-handed content filtering.

Hermes 4 enables users to choose between fast reasoning and more detailed, step-by-step reasoning, with the model’s entire thought process fully transparent.

In maths benchmarks, Hermes 4 405B delivered top-tier results: 96.3% on MATH-500 and 81.9% on AIME’24. Unlike Claude or ChatGPT, which often refuse to answer complex or sensitive queries, Hermes 4 responds two to three times more frequently. Developers can also configure their own censorship levels when deploying the model.

The full code is available here: https://huggingface.co/collections/NousResearch/hermes-4-collection-68a731bfd452e20816725728

Hermes 4 outperforms OpenAI models with minimal content restrictions

MAI-Voice-1, a speech generation model, is already powering news updates and weather reports in Copilot Daily. Anyone can test it by generating custom audio here https://copilot.microsoft.com/labs/audio-expression

MAI-1-Preview is a text-based MoE model which is currently only available on LMArena. During benchmarking, it ranked 13th alongside models such as Grok-3, GPT-4.1 and Gemini 2.5 Flash.

Unlike earlier models, such as those in the Phi series from Microsoft Research, these are the first models to be developed directly within Microsoft AI.

Microsoft AI unveils its first independently developed models

OpenAI has released the Realtime API, which is now open to all developers. This new feature enables developers to build voice-driven agents with support for MCP, image uploads and SIP-based calls. The API also introduces two new voices to facilitate more natural interactions.

Alongside this, OpenAI unveiled GPT-Realtime, its most advanced speech model to date. According to the company, it follows developer instructions more accurately and handles complex tasks with greater reliability.

Developers can try GPT-Realtime here (payment method required): https://platform.openai.com/audio/realtime

OpenAI launches real-time API and new voice model

Google turns Gemini into a Photoshop alternative

![](https://m.stacker.news/106002)
I just asked: "Clean up the land plot."

Google has upgraded Gemini 2.5 Flash to include native image editing capabilities, positioning it as a direct rival to professional design tools. The model can now generate visuals and seamlessly combine and edit existing images.

After weeks of testing under the codename 'nano banana' on LMArena, Gemini consistently ranked first in blind comparisons, surpassing GPT Images and Flux Kontext.

Highlights:
- Preserves the original image quality during edits.
- Merges objects from different photos, generates new angles and replaces backgrounds.
- Supports step-by-step editing, enabling users to gradually refine results.

These new features are available for free in the Gemini app https://gemini.google.com/app and Google AI Studio https://aistudio.google.com/prompts/new_chat

This reasoning AI model is designed for fast coding. While the developers acknowledge its high speed and ability to find bugs faster than competing models, they also caution that if queries are too vague, the model will offer overly simplified solutions.

Grok Code Fast 1 has 314 billion parameters, is built on the Mixture of Experts (MoE) architecture and can process up to 262,000 tokens of context.

Access:
Free on GitHub Copilot until 2 September, then available to Pro, Pro+, Business and Enterprise subscribers.
- a week of free testing in Cursor;
- unlimited in Windsurf for Pro and Teams subscribers.

It costs $0.20 for one million input tokens and $1.50 for one million weekend tokens.

https://docs.x.ai/docs/models/grok-code-fast-1

xAI releases Grok Code Fast 1 model for developers

With 1,500 agents, there isn't a single decisive dating agent? Either there is no market, in which case we are doing the wrong thing, or we are the first to think of solving such an important problem not through endless swiping, but by solving the problems of trust and arranging the first meeting as quickly as possible.

P.S. We are also raising US$25k in angel investments to help us beat the competition. Sitch (YC-backed, 2025) - does not guarantee meetings. Neither does eHarmony (Total Funding: $113M), which also uses 'AI matchmaking'.

1.5k+ AI Agents list

GPT-5 has just been released, yet OpenAI is already working on GPT-6.

The main focus is on memory. ChatGPT should not only respond, but also remember the user's preferences and habits, adapt to them and allow users to customise the bot's tone and personality.

Altman talks about GPT-6

We must build AI for people; not to be a person

According to The Business Insider, American AI IT specialists throw orgies, use drugs and build bunkers. These specialists have lost motivation to live because they believe that 'the world will be destroyed by superintelligence anyway'. A former OpenAI researcher has stopped saving for retirement because he believes that humanity will perish by 2035. His colleagues throw lavish parties in their spare time and spend money recklessly, as it won't be needed in an apocalyptic future.

https://www.businessinsider.com/ai-apocalypse-super-preppers-smart-hot-drugs-bunkers-economic-crash-2025-8

Google has opened free access to Veo 3 until the end of the weekend!

Google has opened free access to Veo3

Alibaba has introduced Qoder, a new agentic coding platform that aims to handle the entire development cycle, from writing code to testing and creating the final build.

Qoder supports two operating modes:
1/ In Agent Mode, the developer supervises each step, acting as a pair programming assistant.
2/ Quest Mode is a fully autonomous agent that generates, tests and delivers production-ready code once the task has been described.

Key features include:
- Deep search and understanding of large codebases, including architecture and design patterns;
- Smart suggestions and auto-generated documentation.
- Long-term memory that adapts to a team’s past decisions and coding style.
- Automatic selection of the most suitable AI model (Claude, Gemini, GPT, etc.) for each task.

Qoder is now available in public preview and is free to try at qoder.com.