Zhipu's largest long position holder failed to withstand the price downtrend, exiting the market with a $2.71 million loss.

Beating AI News – U.S. AI firm Inception has launched its diffusion language model, Mercury 3. Unlike standard large language models (LLMs) that generate tokens one by one, Mercury 3 produces batches of text in parallel and iteratively refines them. Inception notes the new model has over 40 times more parameters than the original Mercury, yet still delivers 1,068 tokens per second on NVIDIA GPUs, nearly matching the 1,107 tokens per second of its predecessor, Mercury 2.5. Its context window has been expanded from 260,000 to 1 million tokens. In published self-tests, Mercury 3’s overall performance at medium inference settings is slightly higher than GPT-6 Luna and Claude Haiku 5.5, with generation speeds 4 to 8 times faster than both. On the DeepSWE coding benchmark, Mercury 3 scores 46.3%, lagging behind Haiku 5.5’s 51.6%; however, it leads with 51.1% on SWE-Bench Atlas, versus the rival’s 32.3%. Inception’s co-founder acknowledged on Hacker News that the new model cannot yet replace Opus or Fable as a primary agent, but can already handle subtasks like code writing. Mercury 3 is now accessible via Inception API, OpenRouter, Baseten, and Vercel. During the launch period, it is priced at 50% off: $0.10 per million input tokens and $0.35 per million output tokens, compared to regular rates of $0.20 and $0.70 respectively. The company is offering 100 million free tokens for trial use. However, the token speed advertised by Inception may not be attainable for regular users. OpenRouter data shows Mercury 3’s median output speed is approximately 180 tokens per second, just one-sixth of the official benchmark speed.

According to official announcements from Bitget, the platform has announced the acquisition of GMO-Z.com Cryptonomics (Thailand) Co., Ltd., a digital asset exchange and broker headquartered in Bangkok. The deal covers all shares of the target company held by shareholders including GMO Internet and its related subsidiaries. GMO-Z.com Cryptonomics (Thailand) holds licenses for digital asset exchange and brokerage services in Thailand. Leveraging its existing compliance foundation and local market expertise, the acquisition will strengthen Bitget’s localized operations in Thailand and further refine its strategic layout in the Southeast Asian market. Bitget CEO Gracy Chen stated that Thailand is a key digital asset market in Southeast Asia, noting that this acquisition will help Bitget establish a localized operational system, gradually expanding the application of its global trading services in the Thai market while complying with local regulatory requirements.

According to official announcements, Bitunix has announced it will raise the size of its Care Fund from $30 million to $50 million, while simultaneously publishing its on-chain address. Users can check the fund’s real-time holdings at any time via a blockchain explorer, ensuring full transparency and verifiability of the fund reserves. The Care Fund, a core component of the platform’s risk protection mechanism, is designed to safeguard user assets during extreme market conditions. This expansion further strengthens the platform’s ability to respond to unexpected risks. The move is a concrete demonstration of Bitunix’s Ultra Trust commitment; the platform will continue to build user trust through transparent, verifiable practices, offering a safer, more reliable trading environment for global traders.

Beating AI News: Ant Group has open-sourced its Ming-3.0 voice model series, spanning three core domains—speech synthesis, audio understanding, and speech recognition—with a total of five models: 2 TTS, 2 Audio, and 1 ASR, all released under the MIT license. Developers can combine these models to build functions like voice input, content analysis, and voice responses. The standout is the 500-million-parameter Ming-3.0-tts-0.5B. Users can use natural language to command the AI to adjust tone (e.g., sounding angry), emphasize specific words, or modify speech rhythm. It also supports voice cloning, custom tone design, and inserting 14 sound tags (such as soft laughs and sighs) at designated positions. In Artificial Analysis’ same-voice blind listening test, the model ranks first among open-source models with an Elo score of 1069. The audio understanding model, Ming-3.0-audio-flash, is far larger, with 126 billion total parameters and ~6.5 billion active parameters per activation. It can directly process meeting recordings to answer questions, summarize consensus, disagreements, and action items without generating a separate transcript first. Another model, Ming-3.0-asr-tiny, specializes in speech recognition, supporting input of names, place names, industry terms, and conversation context to aid dialect and rare word recognition. Across four speech recognition tests conducted by Ant Group, its average error rate is 4.41%.

According to market data from BIT (bit.com), U.S. crypto-related stocks edged lower in pre-market trading, with the following declines among individual names: SBET down 0.8%, BMNR down 1.67%, CRCL down 1.07%, MSTR down 2%, COIN down 1.31%, and ASST down 1.37%

Per Lookonchain monitoring, Drift Protocol has resumed activity after two months of inactivity, transferring ETH to Tornado Cash. Over the past two hours, the attacker has deposited 25,700 ETH (worth roughly $61.56 million) into Tornado Cash, and currently holds 81,864 ETH (valued at approximately $195 million).