7*24

Hot topic

Market

Activities

On June 7, OpenAI is preparing for the largest overhaul of ChatGPT since its launch. The company plans to transform ChatGPT into a 'super app' that combines coding tools and AI capabilities, adding products that executives believe can generate more revenue. (Financial Times)

[OpenAI Prepares Largest Overhaul of ChatGPT Since Launch]On June 7, OpenAI is preparing for the largest overhaul of ChatGPT since its launch. The company plans to transform ChatGPT into a 'super app' that combines coding tools and AI capabilities, adding products that executives believe can generate more revenue. (Financial Times)

OpenAI Prepares Largest Overhaul of ChatGPT Since Launch

According to monitoring by Dongcha Beating, Xiaohongshu's hi lab has open-sourced the end-to-end autoregressive text-to-speech (TTS) model dots.tts, which has 2 billion parameters, and has publicly released the complete inference and fine-tuning code under the Apache 2.0 license. The released weights include a base pre-trained version, a self-correcting alignment (SCA) fine-tuned version, and a low-latency inference distilled version. Unlike traditional TTS architectures that rely on discrete audio codec tokens (such as VALL-E, CosyVoice, ChatTTS, etc.), dots.tts implements a fully continuous, end-to-end autoregressive flow matching architecture that does not use any discrete tokens throughout the pipeline. Dots.tts combines continuous features extracted from AudioVAE at a 48 kHz sampling rate with a semantic encoder, a backbone language model (initialized from Qwen2.5-1.5B-Base, directly processing BPE text without the need for pinyin input), and an autoregressive flow matching acoustic head to predict continuous latent variables, which are reconstructed into audio by the generator. By directly predicting continuous features, dots.tts avoids the quality loss caused by discrete quantization, preserving pronunciation details, timbre similarity, and emotional expressiveness. Dots.tts is pre-trained on approximately 1.5 million hours of speech data. In the Seed-TTS-Eval evaluation, dots.tts achieved word error rates (WER) of 0.94% / 1.30% / 6.60% on Chinese, English, and Chinese difficult test sets, respectively, with similarity scores (SIM) of 81.0 / 77.1 / 79.5, reaching open-source SOTA levels. In the MiniMax Multilingual benchmark across 24 languages, the average speaker similarity reached 83.9. Xiaohongshu has provided a Gradio experience space on Hugging Face for users to test zero-shot voice cloning online.

[Xiaohongshu Releases Open Source Continuous Autoregressive TTS Model dots.tts Supporting Zero-Shot Voice Cloning]According to monitoring by Dongcha Beating, Xiaohongshu's hi lab has open-sourced the end-to-end autoregressive text-to-speech (TTS) model dots.tts, which has 2 billion parameters, and has publicly released the complete inference and fine-tuning code under the Apache 2.0 license. The released weights include a base pre-trained version, a self-correcting alignment (SCA) fine-tuned version, and a low-latency inference distilled version. Unlike traditional TTS architectures that rely on discrete audio codec tokens (such as VALL-E, CosyVoice, ChatTTS, etc.), dots.tts implements a fully continuous, end-to-end autoregressive flow matching architecture that does not use any discrete tokens throughout the pipeline. Dots.tts combines continuous features extracted from AudioVAE at a 48 kHz sampling rate with a semantic encoder, a backbone language model (initialized from Qwen2.5-1.5B-Base, directly processing BPE text without the need for pinyin input), and an autoregressive flow matching acoustic head to predict continuous latent variables, which are reconstructed into audio by the generator. By directly predicting continuous features, dots.tts avoids the quality loss caused by discrete quantization, preserving pronunciation details, timbre similarity, and emotional expressiveness. Dots.tts is pre-trained on approximately 1.5 million hours of speech data. In the Seed-TTS-Eval evaluation, dots.tts achieved word error rates (WER) of 0.94% / 1.30% / 6.60% on Chinese, English, and Chinese difficult test sets, respectively, with similarity scores (SIM) of 81.0 / 77.1 / 79.5, reaching open-source SOTA levels. In the MiniMax Multilingual benchmark across 24 languages, the average speaker similarity reached 83.9. Xiaohongshu has provided a Gradio experience space on Hugging Face for users to test zero-shot voice cloning online.

Xiaohongshu Releases Open Source Continuous Autoregressive TTS Model dots.tts Supporting Zero-Shot Voice Cloning

On June 7, market researcher Financelot indicated that the open-source vector indexing library TurboVec, announced last month, is impacting the high memory demand market, with its effects becoming increasingly apparent. The sharp decline in memory stocks on Friday was attributed to this. Financelot stated, 'Goodbye Micron, SanDisk, Samsung, SK Hynix,' expressing a bearish outlook for the storage sector's performance next week. However, community perspectives suggest that TurboVec's impact on the memory sector is limited; whenever new memory optimizations are announced, there are always claims that the entire semiconductor industry is dead. Google Research introduced the TurboQuant quantization algorithm in March of this year, which was implemented as the open-source vector indexing library TurboVec by independent developer Ryan Codrai in late May. This tool significantly reduces the memory requirements of vector databases (for example, compressing 10 million vectors from float32's 31GB to about 4GB, reducing memory usage by approximately 87%, with savings of up to 16 times depending on dimensions and bit width). It supports fully offline operation, executes efficiently on standard Macs, and offers search speeds that are 12-20% faster on ARM platforms compared to FAISS IndexPQ/FastScan. It is completely open-source and can be integrated with frameworks such as LangChain and LlamaIndex. This means developers can efficiently run local vector searches on standard consumer-grade hardware without relying on expensive GPU clusters or cloud services.

[Outlook: Storage Sector May Continue Downtrend Next Week as Google TurboQuant Algorithm Reduces Memory Demand]On June 7, market researcher Financelot indicated that the open-source vector indexing library TurboVec, announced last month, is impacting the high memory demand market, with its effects becoming increasingly apparent. The sharp decline in memory stocks on Friday was attributed to this. Financelot stated, 'Goodbye Micron, SanDisk, Samsung, SK Hynix,' expressing a bearish outlook for the storage sector's performance next week. However, community perspectives suggest that TurboVec's impact on the memory sector is limited; whenever new memory optimizations are announced, there are always claims that the entire semiconductor industry is dead. Google Research introduced the TurboQuant quantization algorithm in March of this year, which was implemented as the open-source vector indexing library TurboVec by independent developer Ryan Codrai in late May. This tool significantly reduces the memory requirements of vector databases (for example, compressing 10 million vectors from float32's 31GB to about 4GB, reducing memory usage by approximately 87%, with savings of up to 16 times depending on dimensions and bit width). It supports fully offline operation, executes efficiently on standard Macs, and offers search speeds that are 12-20% faster on ARM platforms compared to FAISS IndexPQ/FastScan. It is completely open-source and can be integrated with frameworks such as LangChain and LlamaIndex. This means developers can efficiently run local vector searches on standard consumer-grade hardware without relying on expensive GPU clusters or cloud services.

Outlook: Storage Sector May Continue Downtrend Next Week as Google TurboQuant Algorithm Reduces Memory Demand

On June 7, according to HTX market information, Ethereum has rebounded and surpassed $1600, with a 24-hour increase of 1.8%.

[Ethereum Surges Past $1600, 24-Hour Increase of 1.8%]On June 7, according to HTX market information, Ethereum has rebounded and surpassed $1600, with a 24-hour increase of 1.8%.

Ethereum Surges Past $1600, 24-Hour Increase of 1.8%

On June 7, according to Alternative data, the cryptocurrency Fear and Greed Index today stands at 12 (unchanged from yesterday), indicating that market sentiment continues to be in 'extreme fear'. Note: The fear index ranges from 0 to 100 and includes indicators such as volatility (25%), market trading volume (25%), social media activity (15%), market surveys (15%), Bitcoin's dominance in the market (10%), and Google Trends analysis (10%).

[Crypto Fear Index Reports 12, Market Remains in 'Extreme Fear']On June 7, according to Alternative data, the cryptocurrency Fear and Greed Index today stands at 12 (unchanged from yesterday), indicating that market sentiment continues to be in 'extreme fear'. Note: The fear index ranges from 0 to 100 and includes indicators such as volatility (25%), market trading volume (25%), social media activity (15%), market surveys (15%), Bitcoin's dominance in the market (10%), and Google Trends analysis (10%).

Crypto Fear Index Reports 12, Market Remains in 'Extreme Fear'

According to monitoring by Dongcha Beating, OpenAI is preparing for the largest upgrade of ChatGPT since its launch, seeking new growth engines before going public. OpenAI intends to transform ChatGPT into a 'super app' by integrating programming tools and AI agents, adding several new products that executives believe could generate higher revenue. According to more than a dozen current and former employees, these adjustments are part of OpenAI's overall restructuring plan. OpenAI is reallocating resources to aggressively pursue high-profit enterprise clients and intensify competition with rivals like Anthropic. This adjustment will enhance the status of OpenAI's programming product Codex and increase resource investment, reflecting a growing consensus within the company that the future of artificial intelligence is not merely chatbots that answer questions, but intelligent agents capable of performing tasks for users.

[OpenAI to Significantly Upgrade ChatGPT into a 'Super App', Integrating Programming Tools and AI Agents]According to monitoring by Dongcha Beating, OpenAI is preparing for the largest upgrade of ChatGPT since its launch, seeking new growth engines before going public. OpenAI intends to transform ChatGPT into a 'super app' by integrating programming tools and AI agents, adding several new products that executives believe could generate higher revenue. According to more than a dozen current and former employees, these adjustments are part of OpenAI's overall restructuring plan. OpenAI is reallocating resources to aggressively pursue high-profit enterprise clients and intensify competition with rivals like Anthropic. This adjustment will enhance the status of OpenAI's programming product Codex and increase resource investment, reflecting a growing consensus within the company that the future of artificial intelligence is not merely chatbots that answer questions, but intelligent agents capable of performing tasks for users.

OpenAI to Significantly Upgrade ChatGPT into a 'Super App', Integrating Programming Tools and AI Agents

According to monitoring by Beating, AI startup Boson AI has released the weights for its autoregressive text-to-speech (TTS) model Higgs Audio v3 TTS. The model is built on the Qwen3-4B foundation with approximately 4 billion parameters, specifically optimized for streaming interactions of real-time voice agents. It supports streaming synthesis even before the text is fully generated, reducing latency in real-time voice conversations. Higgs Audio v3 TTS supports over 100 languages and dialects, achieving a single-digit word error rate on test sets such as Seed-TTS, CV3, and MiniMax-Multilingual. The model supports zero-shot voice cloning and allows for the direct embedding of over 20 emotions and various inline control tags (including tone, speech rate, pitch, pauses, as well as effects like coughing, sighing, and laughter) in the input text for highly controllable vocal expressiveness. Boson AI has collaborated with the LMSYS team to optimize the end-to-end serving performance of Higgs Audio v3 TTS on the SGLang-Omni inference framework. Testing on an H100 GPU achieved a single concurrent real-time factor (RTF) of 0.147. The weights have been made publicly available on Hugging Face under a non-commercial research license.

[Boson AI Releases Open Source 4B Audio Model Higgs Audio v3 with Streaming Emotion Control]According to monitoring by Beating, AI startup Boson AI has released the weights for its autoregressive text-to-speech (TTS) model Higgs Audio v3 TTS. The model is built on the Qwen3-4B foundation with approximately 4 billion parameters, specifically optimized for streaming interactions of real-time voice agents. It supports streaming synthesis even before the text is fully generated, reducing latency in real-time voice conversations. Higgs Audio v3 TTS supports over 100 languages and dialects, achieving a single-digit word error rate on test sets such as Seed-TTS, CV3, and MiniMax-Multilingual. The model supports zero-shot voice cloning and allows for the direct embedding of over 20 emotions and various inline control tags (including tone, speech rate, pitch, pauses, as well as effects like coughing, sighing, and laughter) in the input text for highly controllable vocal expressiveness. Boson AI has collaborated with the LMSYS team to optimize the end-to-end serving performance of Higgs Audio v3 TTS on the SGLang-Omni inference framework. Testing on an H100 GPU achieved a single concurrent real-time factor (RTF) of 0.147. The weights have been made publicly available on Hugging Face under a non-commercial research license.

Boson AI Releases Open Source 4B Audio Model Higgs Audio v3 with Streaming Emotion Control

According to monitoring by Dongcha Beating, JetBrains has open-sourced the 12B code mixed expert model Mellum-2, aimed at software engineering. Mellum-2 has a total parameter count of 12 billion, with only 2.5 billion parameters activated per token to control inference overhead. Currently, all weights have been released on Hugging Face under the Apache 2.0 license. The architecture of Mellum-2 introduces a multi-token prediction (MTP) module. During the pre-training phase, the MTP module is trained with auxiliary objectives to optimize representations. During inference, the MTP head acts as a built-in draft model for speculative sampling to accelerate generation, eliminating the need for an external independent small model, and can be directly stripped before production deployment. Mellum-2 employs a grouped query attention (GQA) mechanism with 32 query heads and 4 KV heads, and applies a sliding window attention (SWA) mechanism with a window size of 1024 in three-quarters of its layers, controlling KV cache overhead while achieving a 128K context window through layer-selective YaRN expansion. JetBrains positions Mellum-2 as a focal model in local IDEs or agent workflows, specifically designed for low-latency scenarios such as routing, retrieval, and sub-task validation, rather than directly replacing large-scale frontier models. Mellum-2 is available in three versions: Base, Instruct, and Thinking. The Thinking version, trained with rule-based verification reinforcement learning (RLVR), can display explicit reasoning chains before outputting final answers. In benchmark tests, the Base version scored 41.46% on HumanEval and achieved a 70.87% accuracy on MMLU. The Thinking version scored 69.9% on LiveCodeBench and 58.4% on AIME.

[JetBrains Open Sources Mellum-2 Code Model: Built-in MTP Head Supports Speculative Decoding Acceleration]According to monitoring by Dongcha Beating, JetBrains has open-sourced the 12B code mixed expert model Mellum-2, aimed at software engineering. Mellum-2 has a total parameter count of 12 billion, with only 2.5 billion parameters activated per token to control inference overhead. Currently, all weights have been released on Hugging Face under the Apache 2.0 license. The architecture of Mellum-2 introduces a multi-token prediction (MTP) module. During the pre-training phase, the MTP module is trained with auxiliary objectives to optimize representations. During inference, the MTP head acts as a built-in draft model for speculative sampling to accelerate generation, eliminating the need for an external independent small model, and can be directly stripped before production deployment. Mellum-2 employs a grouped query attention (GQA) mechanism with 32 query heads and 4 KV heads, and applies a sliding window attention (SWA) mechanism with a window size of 1024 in three-quarters of its layers, controlling KV cache overhead while achieving a 128K context window through layer-selective YaRN expansion. JetBrains positions Mellum-2 as a focal model in local IDEs or agent workflows, specifically designed for low-latency scenarios such as routing, retrieval, and sub-task validation, rather than directly replacing large-scale frontier models. Mellum-2 is available in three versions: Base, Instruct, and Thinking. The Thinking version, trained with rule-based verification reinforcement learning (RLVR), can display explicit reasoning chains before outputting final answers. In benchmark tests, the Base version scored 41.46% on HumanEval and achieved a 70.87% accuracy on MMLU. The Thinking version scored 69.9% on LiveCodeBench and 58.4% on AIME.

JetBrains Open Sources Mellum-2 Code Model: Built-in MTP Head Supports Speculative Decoding Acceleration

According to monitoring by Dongcha Beating, Clive Chan, an early member of OpenAI's chip development team and the company's second hardware employee, has announced his departure from OpenAI to join Anthropic this week, where he will work on custom chips and hardware-software co-design. Chan expressed pride in witnessing the team's growth into a leading chip design force over the past two and a half years. While he could not disclose further details about the chips, he noted that, based on the collaboration previously announced between OpenAI and Broadcom, the custom chip project is scheduled to enter production in the second half of 2026. The reason for Chan's departure was the irresistible allure of 'climbing new heights from scratch.' Before joining OpenAI, Chan was involved in the development of the Dojo supercomputer at Tesla.

[OpenAI Hardware Employee Clive Chan Joins Anthropic for Custom Chip Development]According to monitoring by Dongcha Beating, Clive Chan, an early member of OpenAI's chip development team and the company's second hardware employee, has announced his departure from OpenAI to join Anthropic this week, where he will work on custom chips and hardware-software co-design. Chan expressed pride in witnessing the team's growth into a leading chip design force over the past two and a half years. While he could not disclose further details about the chips, he noted that, based on the collaboration previously announced between OpenAI and Broadcom, the custom chip project is scheduled to enter production in the second half of 2026. The reason for Chan's departure was the irresistible allure of 'climbing new heights from scratch.' Before joining OpenAI, Chan was involved in the development of the Dojo supercomputer at Tesla.

OpenAI Hardware Employee Clive Chan Joins Anthropic for Custom Chip Development

The Fed's Bostic said: It's too early to consider a rate cut, and in my forecast, there will be no recession at this time.

Detail

【Fed's Bostic: It's Too Early to Consider Rate Cuts】The Fed's Bostic said: It's too early to consider a rate cut, and in my forecast, there will be no recession at this time.

Fed's Bostic: It's Too Early to Consider Rate Cuts

A $300 Used Mining Rig Mined a $230,000 Bitcoin Block

The End of the 'Bitcoin Hoarding Myth': MicroStrategy Sells BTC for the First Time in 3 Years

Selling 32 BTC to Pay Dividends: Strategy Breaks the 'Never Sell' Vow

Another Bitcoin Buy Ahead? Michael Saylor’s Latest Post Fuels Rumors

Strategy Ends Bitcoin Buying Streak With Rare Sale—How Much Did It Sell?

AMD Stock Price Forecast 2026 to 2030: Wall Street Signals a Major $657 Target!

Over $1 Billion in Cryptocurrencies Forced to Liquidate, Bitcoin Plunges 6% Below $67,000

How To Avoid The Major Trap That Bitcoin Is Setting Up For Traders

Bitcoin Whales Most Active In Six Weeks As BTC Drops Under $67,000

Bulls Approach the Ceiling

Article

Bitwise CEO: A New Cycle Has Begun, the Crypto Market is Experiencing 'Renewal' This Year

Fed's Bostic: It's Too Early to Consider Rate Cuts

Penafian：Konten diatas tidak merepresentasikan posisi HTX.，HTX tidak memberikan rekomendasi perdagangan apa pun.。

Artikel Terkait

Semua Komentar0TerkiniHangat

Topik Trending

Artikel Terkait