跳到正文
#模型发布
10月6日周二
10月5日周一
10月4日周日
10月3日周六
10月2日周五
  2. Ai2 / Allen Institute for AI62

    Ai2 开源 8B 科学报告生成模型 AstaBrief

    Ai2 开源 AstaBrief 8B，一个基于 Qwen3-8B、将研究问题和检索文献片段转化为带引用报告的科学报告生成模型，现已在 Asta 的 Generate a report 功能中作为 Fast mode 上线，并连同训练数据开放下载。

    推荐理由：原文给出训练数据构成、过滤方法与速度成本对比，读者可据此判断开源科学报告模型的具体做法是否可迁移。

  4. IT之家·人工智能67

    Black Forest Labs 发布生图模型 FLUX 3 Image：支持 4K 生成与元素精准排布

    德国 AI 公司 Black Forest Labs 于 10 月 2 日发布图像生成模型 FLUX 3 Image，支持生成最高 4K 分辨率图片。用户可在 0–1000 坐标网格上通过 ID、描述文本和边界框精准排布画面元素，单次生成最多融入 10 张参考图像，并支持像素级精确多轮编辑。该模型目前以付费服务提供，开放模型版本将于数周内公开。

  12. Rohan Paul62

    Tavus 发布首个 Human Interaction Model（HIM）Griffin，称其为首个通过视频图灵测试的模型。48% 与其实时视频对话的人误以为对方是真人，此前系统通过率低于 3%，并位列 NVIDIA 全双工 AI 视频基准第一。作者 Rohan Paul 转引 Tavus 官方发布并补充演示视频中的表现。

    引用Tavus@tavus

    Introducing Griffin, the first model to pass the video Turing test. 48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video. It’s the first Human Interaction Model (HIM).

  14. fal54

    Tavus 发布 Griffin，称其为首个通过视频图灵测试的模型，属于 Human Interaction Model（HIM）。与 Griffin 实时对话的受访者中 48% 认为它是真人，此前系统通过率低于 3%，并在 NVIDIA 的全双工 AI 视频基准上排名第一。fal 转发该发布并祝贺 Tavus 团队。

    引用Tavus@tavus

    Introducing Griffin, the first model to pass the video Turing test. 48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video. It’s the first Human Interaction Model (HIM).

  16. Rohan Paul64

    Tavus 发布 Griffin，称其为首个通过视频图灵测试的人类交互模型（HIM），48% 与之实时对话的人认为它是真人，此前系统通过率低于 3%。Griffin 在 NVIDIA 全双工 AI 视频基准上得 3.83 分（真人为 3.92），次优系统为 2.80 分；它将听、想、说和面部动画整合在单一系统中，可在对话者仍在说话时读取表情、语气并实时反应。

    引用Tavus@tavus

    Introducing Griffin, the first model to pass the video Turing test. 48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video. It’s the first Human Interaction Model (HIM).

  20. Mustafa Suleyman64

    Microsoft AI 发布流式语音转写模型 MAI-Transcribe-2-Streaming，Mustafa Suleyman 称其为全球最准确的实时转写模型，比 ElevenLabs 快 55%、便宜 60%。引用的 Artificial Analysis 数据显示，该模型以 2.5% WER、语音结束后 0.13 秒出稿在 AA-WER Streaming 38 个模型中排名第一，超过此前第一的 Grok Voice Transcribe 2.0（2.7%、0.49s），流式定价为每小时音频 $0.54，非流式为 $0.10 每小时。

    引用Artificial Analysis@ArtificialAnlys

    Microsoft AI has released MAI-Transcribe-2-Streaming, taking the #1 spot for Final Transcript accuracy and First Partial Transcript accuracy on AA-WER Streaming with 2.5% WER at 0.13s after end of speech MAI-Transcribe-2-Streaming is @MicrosoftAI's new streaming Speech to Text model, joining the non-streaming MAI-Transcribe-2. It leads streaming Final Transcript WER at 2.5%, ahead of the previous #1, SpaceXAI's Grok Voice Transcribe 2.0 at 2.7%, and returns that transcript in 0.13s rather than 0.49s. It is available for streaming transcription at $0.54 per hour of audio, at the higher end of pricing among the leading streaming models. Key takeaways ➤ Final Transcript: MAI-Transcribe-2-Streaming achieves 2.5% WER at 0.13s after end of speech, ranking #1 of 38 models. It is more accurate and faster than Grok Voice Transcribe 2.0 at 2.7% and 0.49s, Muse Voice Transcribe at 3.1% and 0.16s, and ElevenLabs Scribe v2 Realtime at 3.6% and 0.14s. It is also more accurate, though slightly slower, than Cartesia Ink Preview (external endpoints) at 3.1% and 0.11s ➤ First Partial Transcript: The model achieves 2.5% WER at 0.12s, ahead of Grok Voice Transcribe 2.0 at 3.4% and 0.49s, and ahead of Muse Voice Transcribe and ElevenLabs Scribe v2 Realtime on accuracy, both at 3.6%, and slightly faster than both at 0.12s versus 0.13s. It is more accurate but slower than Cartesia Ink-2 (external endpoints) at 4.0% and 0.07s ➤ Price: MAI-Transcribe-2-Streaming costs $0.54 per hour for streaming, or $9.00 per 1,000 minutes. This puts it level with Gemini 3.5 Transcribe Live at $9, above the $6.50 charged for ElevenLabs Scribe v2 Realtime and Deepgram Flux, more than twice Cartesia Ink-2 at $4, and three times Muse Voice Transcribe at $3. Non-streaming transcription costs $0.10 per hour, or $1.67 per 1,000 minutes Congrats to the @MicrosoftAI team on the launch! See more details below

10月1日周四
  5. IT之家·人工智能65

    谷歌逐步推出 Gemini 4 Argon 旗舰模型，内部对编码实战表现存在质疑

    据彭博社报道，谷歌开始向小批网络安全合作伙伴逐步推出旗舰模型 Gemini 4 Argon，后续将优先面向付费订阅用户扩大开放。谷歌称该模型在多项基准测试中取得靠前成绩，安全测试成绩超过 OpenAI 的 Astra；但知情人士称其部分代码任务表现不佳、不擅前端设计，谷歌否认这一说法，DeepMind 负责人卡武库奥卢表示对性能感到鼓舞。