Daily AI Briefing: Zhipu Open-Sources GLM-4.7; MiniMax and Zhipu Pass HKEX Hearings

Dec 23 · Tuesday

GLM-4.7 Launched and Open-Sourced: Stronger Coding

Zhipu AI has officially launched and open-sourced its new-generation large model GLM-4.7, which achieves significant improvements in coding, reasoning, and agent capabilities. In multiple mainstream benchmarks, its programming capability reaches the leading level among open-source models, achieving a SOTA score of 73.8% on SWE-bench Verified and 87.4 on the tool-calling task τ²-Bench, surpassing Claude Sonnet 4.5. Source: Zhipu

Dec 22 · Monday

MiniMax Passes HKEX Hearing, Poised to Become the "World's First AGI Stock"

MiniMax has passed the Hong Kong Stock Exchange hearing and disclosed its prospectus, sprinting toward an IPO. Founded in early 2022, MiniMax would be the fastest AI company globally from founding to IPO if it successfully lists. In 2024, MiniMax's revenue grew 782% year over year to $30.5 million, with a net loss of $465 million. Its monetization methods include subscription services, token-based in-app purchases, online marketing services, and usage-based enterprise API services. Source: Yicai

Zhipu Sprinting for Hong Kong Stock Exchange's "First Large Model Stock," Valued at 24.4 Billion

Zhipu has passed the Hong Kong Stock Exchange hearing, sprinting to become the Hong Kong Stock Exchange's "first large model stock." Founded in June 2019, Zhipu is China's largest independent general large model developer and China's second-largest overall general large model developer. Its MaaS platform offers four types of models—language, multimodal, agent, and code—and supports both on-premises and cloud deployment. In November 2025, Zhipu's average daily token consumption reached 4.2 trillion, and it already provides support for approximately 80 million devices. Source: Zhidongxi

NVIDIA's New Model NitroGen Can Play Almost Any Game

NVIDIA has open-sourced the general game AI model NitroGen, which can directly generate gamepad operation signals by watching game footage and supports zero-shot play across more than 1,000 games of different types. The model is trained on a dataset of 40,000 hours of game videos and adopts the GR00T architecture designed for robots, possessing strong cross-game generalization capability. Source: Machine Heart

Dec 19 · Friday

Coze Coding Opens Free Public Beta

At the Volcano Engine Force conference, the Coze development platform was officially upgraded to "Coze Coding" and opened a free public beta. This upgrade marks the platform's transformation from an AI application-building tool into a platform that empowers developers' creativity, introducing a new Vibe Coding development paradigm. Users only need to describe business requirements in natural language to automatically generate agents, workflows, and cross-platform applications. Source: Coze

OpenAI's Strongest Code Model GPT-5.2-Codex Launched

OpenAI has released GPT-5.2-Codex, its most advanced agentic coding model to date, achieving major breakthroughs in complex software engineering tasks. The new version significantly improves instruction-following capability, long-context understanding, and cybersecurity protection performance, reaching state-of-the-art levels on benchmarks such as SWE-Bench Pro and Terminal-Bench 2.0. Its unique context compression technology can efficiently handle large codebases and support complex refactoring and migration tasks. Source: Machine Heart

VTP: MiniMax Hailuo Video Team's First Open-Source Release!

The MiniMax video team has open-sourced VTP, a key component of its visual generation model, and is the first to propose the scaling property of tokenizers. The technology integrates multiple representation learning methods to establish a connection between latent learnability and general representation learning, significantly improving the performance of downstream generation systems. Source: MiniMax

Apple Open-Sources SHARP Model, Bringing Photos to 3D "Life" with AI in 1 Second

Apple has announced the open-sourcing of its new AI model SHARP, which can convert a single 2D photo into a realistic 3D scene within one second. Compared with traditional solutions, processing speed is improved by three orders of magnitude, achieving a near-real-time 3D conversion experience. The model is based on advanced 3D Gaussian splatting technology and can predict the position and appearance of millions of 3D Gaussian spheres through a single neural network feedforward pass. Source: IT Home

Dec 18 · Thursday

Google Releases High-Speed, Low-Cost Model Gemini 3 Flash

Google has officially released the high-speed, low-cost model Gemini 3 Flash, which performs excellently in multiple benchmarks, surpassing its predecessor Gemini 2.5 Pro and approaching frontier models such as Gemini 3 Pro and GPT-5.2 on some metrics. The new model features frontier-level reasoning and multimodal capabilities, with response speed increased by 3 times while significantly reducing costs. Source: Machine Heart

Meituan Releases LongCat-Video-Avatar, Achieving Open-Source SOTA-Level Realism

Meituan has released and open-sourced the SOTA-level virtual human video generation model LongCat-Video-Avatar, which achieves significant breakthroughs in three dimensions: motion realism, long-video stability, and identity consistency. The model natively supports functions such as joint audio-video generation and video continuation. Through innovative cross-clip latent space stitching technology, it effectively solves the image quality degradation problem in long-video generation and can stably generate 5-minute high-quality videos. Source: LongCat

General-purpose Agent Model Seed1.8 Officially Released

ByteDance has officially released the general-purpose Agent model Seed1.8. The model integrates diverse capabilities including search, code generation, and graphical user interface interaction, supports image-text input, and features native visual understanding. It performs excellently on multiple benchmarks, achieving a score of 61.9 on OSWorld for GUI Agent tasks and 67.6 on BrowseComp-en for search tasks, both at industry-leading levels. Source: ByteDance Seed

Bigger and Stronger! Limited-Time Free! StepFun GUI Agent Fully Upgraded

StepFun has announced a comprehensive upgrade to its GUI Agent model Step-GUI. The new version supports over 200 task scenarios and features longer reasoning steps, stronger semantic understanding, and generalization capabilities. The model can be used across multiple platforms including phones, PCs, and cars, and enables rapid deployment in as little as 10 minutes. The company has opened an API for free use and publicly released a technical report and the GUI-MCP protocol. Source: StepFun

Meta Releases SAM Audio, a New Breakthrough in Multimodal Audio Separation

Meta has released SAM Audio, the first unified multimodal audio separation model, which supports precisely extracting target sounds from complex audio via text, visual, or time-segment prompts. Based on a perceptual encoder audiovisual technology, the model allows users to click on objects in a video to separate the corresponding audio, or input text instructions such as "dog barking" to filter out noise. Source: IT Home

Valuation Reaches 46.5 Billion! Star AI Coding Unicorn Lovable Secures New Funding

Swedish AI coding startup Lovable recently completed a new round of funding, with its valuation soaring to 6.6 billion USD, more than tripling in half a year and making it one of Europe's most valuable startups. This round attracted participation from well-known venture capital firms including Accel, marking intensifying competition in the AI coding tools track. Source: Zhidongxi

AIsphere and Alibaba Cloud Sign Comprehensive Cooperation Agreement. The two parties will establish deep collaboration in the field of AI video generation. Alibaba Cloud will provide AIsphere with full-stack AI support, covering infrastructure and large model services, to help with the global deployment of its self-developed video generation model PixVerse. Source: AIsphere

December 17 · Wednesday

Jimeng AI Web Version Fully Upgraded to Create a One-Stop "AI Studio". Core updates include: the launch of the "Video 3.5 Pro" model based on Seedance 1.5 Pro, supporting joint audio-video generation and dialect lip-sync matching; Smart Multi-Frame 2.0 enabling "video + video" stitching and fine-grained clip editing, supporting one-shot continuity with up to 20 materials; and the Image 4.1 and 4.5 models strengthening graphic design capabilities and portrait aesthetics, respectively. Source: Jimeng AI

OpenAI's Strongest AI Image Generation Model GPT Image 1.5 Debuts

OpenAI has released its new flagship image generation model GPT Image 1.5. The model achieves major breakthroughs in prompt adherence, detail preservation, and text rendering, and can precisely generate or edit images while maintaining core features of the original image such as lighting and composition. Source: IT Home

Xiaomi Officially Releases and Open-Sources New Model MiMo-V2-Flash

Xiaomi has officially released the open-source large model MiMo-V2-Flash. The model adopts a MoE architecture with 309 billion total parameters and 15 billion active parameters. It performs excellently on multiple benchmarks, with particularly outstanding coding capabilities, surpassing all open-source models on the SWE-bench test with a 73.4% resolution rate. Through technical innovations such as hybrid sliding window attention and multi-token prediction, the model achieves a high inference speed of 150 tokens/second and extremely low inference costs. Source: APPSO

Tencent Hunyuan World Model 1.5 Released, China's First to Open Real-Time Experience

Tencent has officially released the Hunyuan World Model 1.5 (WorldPlay), China's first world model supporting real-time interactive experience. Users can generate their own 3D interactive worlds by inputting text or images, and freely explore them via keyboard, mouse, or gamepad. The model features 24FPS real-time generation, minute-level 3D geometric consistency, and spatial memory capabilities, keeping the scene consistent upon returning after leaving. Source: Tencent Hunyuan

Seedance 1.5 pro Audio-Video Creation Model Officially Released

ByteDance has released its new-generation audio-video creation model Seedance 1.5 pro, achieving joint audio-video generation. The model supports text-to-audio-video synthesis and image-guided generation, and has made breakthroughs in precise audio-visual synchronization, multilingual and dialect lip-sync matching, cinematic camera movement control, and dynamic narrative performance. Source: ByteDance Seed

December 16 · Tuesday

LiblibAI Launches "Wan 2.6" Video Model, the Chinese Version of SORA2 Is Here!

LiblibAI platform has globally premiered the Tongyi Wanxiang Wan 2.6 video generation model, supporting three core functions: video reference generation, audio-visual synchronization, and intelligent multi-shot scheduling. The model can directly output 15-second 1080P high-definition videos, achieving complete narrative generation without the need for secondary editing. Its video reference generation capability can replicate any character's appearance and voice within 5 seconds, and supports audio-visual synchronization for single/dual performances and multi-person dialogue scenes. Source: LiblibAI

SenseTime Xiaohuanxiong 3.0 Is Here, AI Office Agent Generates High-Quality PPTs with One Click

SenseTime releases AI office agent "Xiaohuanxiong 3.0," achieving three major leaps from generating drafts to delivering finished products. The agent can generate high-quality PPTs with one click, supports editing of all elements including charts and images, and can deliver directly usable presentations. Its "long-chain thinking" capability supports second-level processing of million-scale data volumes and multimodal analysis, and can deeply understand complex tasks. Source: SenseTime

Wan2.6 officially released! Launching the first role-playing feature in China

Alibaba released the Tongyi Wanxiang 2.6 multimodal model, the first video generation system in China to support a role-playing feature. The model supports audio-visual synchronization, multi-shot narrative, and audio-driven generation, and can directly output 15-second high-definition videos. The role-playing feature allows users to upload reference videos, and the AI can replicate a character's appearance and voice to generate single-person or multi-person performance content. In image generation, the model improves artistic style fusion and detail depiction, and supports commercial-grade creation such as mixed text-image layouts and multi-image fusion. Source: Tongyi Wanxiang Wan

NVIDIA releases Nemotron 3 open-source AI model series, throughput 4 times that of the previous generation

English text is being prepared.

Tongyi Bailing speech twins, open-sourced simultaneously!

Alibaba Cloud's Tongyi Bailing open-sources the speech twin models Fun-CosyVoice3 and Fun-ASR. Fun-CosyVoice3 supports voice cloning and emotion control across 9 languages and 18 dialects with just 3 seconds of audio, reduces first-packet latency by 50%, and lowers the error rate for mixed Chinese-English speech by 56.4%. Fun-ASR achieves 93% recognition accuracy in noisy environments, supports mixed speech in 31 languages and dialect recognition, and reduces the first-character latency of streaming recognition to 160 milliseconds. Source: Tongyi Large Models

December 15 · Monday

SenseTime Seko2.0 grandly released, co-produced short drama tops Douyin AI short drama chart at No.1

SenseTime released the Seko 2.0 multi-episode generation agent, the industry's first integrated creation and editing AI tool, designed specifically for short drama creators. This version solves the challenge of character consistency across multiple episodes through SekoIDX technology, and SekoTalk realizes the industry's first precise lip-sync feature for more than 2 people, which can shorten the traditional 3-month production cycle for 50 episodes by 80%-90%. At the same time, it open-sources the LightX2V inference framework, which can generate a 5-second video in 5 seconds on consumer-grade graphics cards, and has completed adaptation for domestic chips. Source: SenseTime

Ant's AI health assistant AQ adopts the Chinese name "Ant Afu"

English text is being prepared.

Google Gemini TTS upgrade: supports 24 languages, more realistic voices

Google's Gemini TTS model: released upgraded versions of the Gemini 2.5 Flash and Pro text-to-speech models, improving expressiveness, speech rate control, and performance in multi-speaker scenarios. The new models provide better control over style, intonation, and pronunciation, can adjust speech rate according to context—such as speeding up when excited and slowing down for emphasis—and can maintain consistent character voices in multilingual dialogue, making conversations more realistic. Source: AI Tool Collection

December 12 · Friday

GPT-5.2 officially released! Saving workers 10 hours per week

OpenAI officially released the GPT-5.2 model, launching three versions: Instant, Thinking, and Pro, targeting routine tasks, complex reasoning, and high-precision needs, respectively. The model performs outstandingly in multiple benchmark tests, reaching or surpassing human expert level for the first time in the GDPval knowledge work task evaluation, and setting a new record of 55.6% in software engineering capability (SWE-Bench Pro). Its hallucination rate is reduced by about 30%, and long-text processing accuracy is close to perfect. Source: APPSO

Zhipu multimodal open-source week concludes: four core video generation technologies open-sourced

At the conclusion of its multimodal open-source week, Zhipu AI announced the open-sourcing of four core video generation technologies: the SCAIL film-grade character animation framework achieves complex pose control; the RealVideo real-time streaming video generation system compresses first-response latency to 2-3 seconds; the Kaleido multi-subject generation framework reaches SOTA level among open-source models; and the SSVAE spectral structured variational autoencoder triples model convergence speed. Source: Zhipu

Google's major open-source release! Deep research Agent achieves SOTA

Google has open-sourced its deep research Agent Gemini Deep Research, which reaches a 46.4% SOTA level on the DeepSearchQA benchmark, with performance comparable to GPT-5 Pro while costing only one-tenth as much. The Agent is based on the Gemini 3 Pro model, possesses multi-step search and complex information integration capabilities, and can automatically generate in-depth research reports. Source: Zhidongxi

Disney officially announces a $1 billion investment in OpenAI, allowing Sora to generate videos of Mickey Mouse and other characters

English text is being prepared.

Thursday, December 11

Zhipu open-sources GLM-TTS: controllable-pronunciation speech synthesis based on multi-reward reinforcement learning

Zhipu AI released and open-sourced the industrial-grade speech synthesis system GLM-TTS. It adopts a two-stage generation paradigm and supports voice timbre replication with only 3 seconds of audio and cloning of multiple dialects. After reinforcement learning optimization, the character error rate (CER) reaches 0.89%, achieving SOTA level among open-source models. Key technical breakthroughs include multi-reward fusion reinforcement learning, fine-grained pronunciation control (Phoneme-in), and the self-developed 2D-Vocos vocoder, significantly improving emotional expression and pronunciation accuracy. Source: Zhipu

ModelBest open-sources VoxCPM 1.5, upgrading speech generation capabilities again

Version 1.5 of VoxCPM has officially been open-sourced and launched. It raises the AudioVAE sampling rate from 16kHz to 44.1kHz, significantly improving the level of detail in voice cloning. Generation efficiency has doubled, requiring only 6.25 tokens to generate 1 second of audio. At the same time, LoRA and full fine-tuning scripts have been added to support deep customization, and model stability has been optimized to reduce audio artifacts. Source: ModelBest

Alibaba releases Qwen3-Omni-Flash full-modal large model

English text is being prepared.

Adobe and ChatGPT team up to launch new image and PDF editing features

English text is being prepared.

Wednesday, December 10

GLM-ASR open-sourced, Zhipu AI Input Method released

Zhipu AI released and open-sourced the GLM-ASR series of speech recognition models, while also launching the desktop "Zhipu AI Input Method." The series includes the leading cloud model GLM-ASR-2512 and the open-source SOTA on-device model GLM-ASR-Nano-2512 with only 1.5B parameters; the latter achieves stronger privacy protection and lower latency while ensuring high accuracy. Source: Zhipu

Mistral AI open-sources a new-generation coding model: Devstral 2

Mistral open-sourced the Devstral 2 series of code large models, including 123B and 24B versions, both supporting a 256K context. The models score 72.2 and 68.0 respectively on SWE-bench Verified, reaching the top level among open-source models, while being significantly smaller in size than similar competitors. The accompanying Vibe CLI tool can integrate model capabilities into the terminal, supporting operations such as scanning project structures, invoking commands, and automatically generating code changes. Source: AI Tool Collection

Generative AI media platform Fal receives another $140 million in funding

Generative AI media platform Fal has secured a new $140 million funding round led by Sequoia Capital, with participation from institutions including Nvidia's venture arm Nventures. The company's valuation has reached $4.5 billion, tripling compared with a few months ago. The company provides developers with hosting services for image, video, and audio AI models, and its clients include enterprises such as Adobe and Shopify. Source: AI Tool Collection

Tuesday, December 9

AutoGLM open-sourced: every phone can become an AI phone

Zhipu AI announced the open-sourcing of the AutoGLM project, enabling AI to truly learn to "use a phone." After 32 months of R&D, the project built a complete Phone Use capability framework, allowing AI to understand the phone interface through vision and complete operations such as tapping and swiping, realizing automated tasks ranging from ordering takeout to batch-processing notifications. To safeguard privacy and security, the system mainly runs in a cloud-based virtual phone environment. Source: Zhipu

Ant Group officially launched Lingguang, a full-modal general-purpose AI assistant, on the web, which users can access directly through a browser. The web version continues the core advantage of "generating a mini-app in natural language in 30 seconds" and achieves synchronization of data and creations with the mobile end. Users can describe their needs on a computer, quickly generate, edit, and manage conversations and flash apps, and continue using and sharing them on mobile. Source: IT Home

Monday, December 8

Zhipu officially launches and open-sources the GLM-4.6V series of multimodal large models

Zhipu AI officially launched and open-sourced the GLM-4.6V series of multimodal large models, including 106B and 9B versions. The models natively support autonomous tool invocation based on visual input and can handle complex tasks such as mixed text-image layouts and image-based shopping. Their 128K long context window can understand documents up to 150 pages long or 1 hour of video content, with improved capabilities in areas such as front-end code replication. Source: AI Tool Collection

Meituan releases the LongCat-Image image generation model, with editing capabilities topping the open-source SOTA

Meituan released and open-sourced the LongCat-Image image generation model, which with 6B parameters approaches larger models in text-to-image and image editing capabilities. Technical highlights include: adopting a homologous architecture and progressive learning to reach open-source SOTA in image editing controllability; significantly improving the coverage and accuracy of Chinese text generation through curriculum learning and character-level encoding; and enhancing image realism by combining data filtering with adversarial training. Source: LongCat

Hunyuan 2.0 goes live: reasoning capability and efficiency leading the industry

Tencent released Hunyuan 2.0, the latest version of its Hunyuan large model, including Think and Instruct versions. The model adopts a mixture-of-experts architecture with 406B total parameters and supports a 256K context window. It performs excellently on complex reasoning tasks such as mathematics, science, and code, with comprehensive capabilities ranking in China's first tier. By optimizing training strategies, the model significantly reduces chain-of-thought length while maintaining high accuracy, leading in reasoning efficiency. Source: Tencent Hunyuan

AI legal tech company Harvey secures new funding (Series F), valued at 56 billion!

U.S. AI legal tech company Harvey announced the completion of a $160 million Series F funding round, reaching a valuation of $8 billion. Founded in 2022 by former Google DeepMind researchers and top law firm lawyers, the company focuses on the "AI + legal" field. Its legal AI platform includes features such as an AI assistant and a document library, and it has partnered with 50 of the top 100 U.S. law firms. Harvey has raised 7 rounds totaling more than $960 million, with participation from OpenAI Startup Fund, Google Ventures, and others multiple times. Source: Zhidongxi

Friday, December 5

Hears clearly, understands well! Doubao speech recognition model 2.0 is here

Volcano Engine released the Doubao speech recognition model 2.0. Based on the Seed mixture-of-experts architecture, the model achieves deep contextual reasoning through PPO reinforcement learning, increasing keyword recall by 20%. It adds multimodal visual recognition capability, which can accurately distinguish easily confused words (such as "hua ji" [slippery chicken] and "hua ji" [funny]) in combination with image content, and supports accurate recognition of 13 languages including Japanese, Korean, and German. Source: Volcano Engine

Kling Digital Human 2.0 officially launched! Not only can it speak, it can also act!

Kling fully launched version 2.0 of its digital human, achieving three core upgrades: fully evolved performance capability, supporting precise control of posture, gestures, and expressions; significantly optimized lip-sync and hand details for enhanced realism; and extended single-generation video duration to 5 minutes, covering long-form content scenarios such as popular science and advertising. Source: Kling AI

Google begins public testing of "Olympiad gold medal-level" reasoning AI model Gemini 3 Deep Think

Google officially began public testing of Gemini 3 Deep Think mode for Gemini Ultra subscribers ($249.99 per month). The mode uses parallel reasoning technology, invoking more computing resources, and achieves breakthroughs in complex math, science, and logic problems: scoring 41% on the Humanity's Last Exam test and 93.8% on the GPQA Diamond test. Source: IT Home

AI video generation product Pollo AI raises $14 million in funding

AI video generation platform Pollo AI recently completed a $14 million first funding round, led by Gaocheng Capital with participation from ZhenFund. Its founder Zhu Chenbiao (A Biao) has no big-tech background and used SEO experience to drive early product growth, surpassing 4 million monthly active users in 7 months. The platform currently has more than 20 million registered users and has achieved break-even. Source: LatePost

Google launches Workspace Studio to empower all employees to build AI agents

Google officially released Workspace Studio, a new-generation automation tool based on the multimodal capabilities of Gemini 3. Users can quickly design, manage, and share AI agents without coding, automating everything from simple tasks to complex workflows. Source: AI Tool Collection

December 4 · Thursday

LiblibAI launches Seedream 4.5, with comprehensive upgrades to consistency preservation and multi-image reference capabilities!

The domestic SOTA image model Seedream 4.5 is officially launched on the LiblibAI platform. Compared with version 4.0, the new model has significantly upgraded capabilities in consistency preservation, precise text/layout rendering, multi-image reference generation, and realistic texture simulation. It is especially skilled at handling localized content such as Chinese posters and marketing materials, and supports the fusion of up to 10+ reference images. Source: LiblibAI

Xiaohongshu acquires "Diandian"

Xiaohongshu has acquired Shanghai Shengsheng Shizhang Technology Co., Ltd. in full, formally incorporating its AI search product "Diandian" into the group system. This acquisition strengthens AI search capabilities and enhances competitiveness in content e-commerce. After the acquisition, Xiaohongshu will accelerate the implementation of its AI strategy, respond to competition from giants such as ByteDance and Alibaba, optimize user behavior paths, and consolidate its content ecosystem. Source: 36Kr

Kling 2.6 fully launched! Hear the visuals, see the sound

Kling 2.6 is fully launched. It is its first "audio-visual simultaneous output" model, capable of generating visuals, natural speech, matching sound effects, and ambient atmosphere simultaneously in a single generation. The model supports two creation paths: text-to-audio-visual and image-to-audio-visual, and is suitable for a variety of applications such as solo monologue, narration, multi-person dialogue, musical performance, and creative scenes. Source: Kling AI

Doubao image creation model Seedream 4.5 released: focused on commercial productivity scenarios

Volcano Engine released the Doubao image creation model Seedream 4.5 and opened public beta. The model has significantly improved subject consistency, instruction following, spatial logic, and aesthetic expression, with a focus on strengthening multi-image combination generation, poster layout, and Logo design capabilities, and can precisely meet the needs of commercial scenarios such as advertising and marketing, e-commerce operations, and film and television production. Source: Volcano Engine

December 3 · Wednesday

Mistral AI releases the Mistral 3 series models, with the entire line returning to Apache 2.0

Mistral AI released a new generation of Mistral 3 series models, with the entire line returning to the Apache 2.0 open-source license. The series includes the 675B-parameter Mistral Large 3 mixture-of-experts model, as well as 14B/8B/3B Ministral 3 edge-side models. Large 3 was trained on 3,000 H200 GPUs, has image understanding capabilities, and ranks second among open-source models on LMArena. Source: Machine Heart

Anthropic acquires popular JavaScript toolchain company Bun

Anthropic announced the acquisition of the popular JavaScript toolchain company Bun, with the transaction amount undisclosed. Bun founder Jarred Sumner chose to join because his tool is widely used by AI programming products such as Claude Code, and because he is optimistic about Anthropic's prospects in the AI programming field. This acquisition deeply integrates Bun's high-performance runtime and bundling technology into Claude Code and future AI tools to improve performance and distribution efficiency. Bun will remain open source and continue to develop independently. Source: InfoQ

Amazon Web Services releases ten major AI updates including the Nova 2 series AI models

At the AWS re:Invent conference, Amazon Web Services released ten major AI updates: the launch of its self-developed AI chip Trainium4 (6x performance improvement) and Trainium3 servers; the second-generation large model Amazon Nova 2 series (Lite, Pro, Sonic, Omni), whose performance surpasses many mainstream models; the Bedrock platform added 18 open-source models, including Chinese models such as Alibaba's and Kimi; and the launch of AgentCore development tools and four major agents (such as autonomous programming and security operations), strengthening enterprise-level AI applications. Source: Zhidongxi

SenseTime releases the NEO architecture, redefining the performance boundaries of multimodal models

SenseTime, together with Nanyang Technological University, released the world's first native multimodal architecture NEO, breaking through the traditional patchwork paradigm of "visual encoder + projector + language model." Through three major innovations—native patch embedding, 3D rotary position encoding, and native multi-head attention—the architecture achieves deep fusion of vision and language. Source: SenseTime

December 2 · Tuesday

LiblibAI launches the "Kling O1" all-around video model, come and experience it

LiblibAI globally premiered the "Kling O1" unified multimodal video model, the first unified model that integrates multiple tasks such as video content addition and deletion, style repainting, and shot extension into one. The model supports up to 7 reference images to maintain content consistency, can intelligently edit videos through natural language instructions, and has video reference generation capabilities. Source: LiblibAI

Runway releases the latest video generation AI model Gen-4.5, codenamed "David"

Runway officially releases the Gen-4.5 video generation model, which entered the Artificial Analysis leaderboard in first place under the anonymous identity "Whisper Thunder (aka) David." Gen-4.5 sets new standards in motion quality, prompt adherence, and visual realism, achieving an ELO score of 1247 and surpassing mainstream models such as Veo 3/3.1, Kling 2.5, and Sora 2 Pro. It features precise physical simulation capabilities, complex scene construction, and cinematic visual quality, and supports multiple control modes. Source: Machine Heart

NVIDIA unveils a reasoning version of VLA: Alpamayo-R1 makes autonomous driving AI better at thinking

NVIDIA releases the reasoning version of its vision-language-action model Alpamayo-R1, giving autonomous driving systems causal reasoning capabilities. The model breaks through the limitations of traditional end-to-end systems in long-tail scenarios, using a "causal chain" dataset to let vehicles understand "why to drive this way" rather than simply executing instructions. The three major innovations include a structured causal annotation system, a diffusion-based trajectory decoder, and a multi-stage training strategy. Source: Machine Heart

PixVerse V5.5 is here! "Director-level" audio-visual synchronization generated with one click

PixVerse releases the PixVerse V5.5 video generation model, achieving China's first synchronized storyboard and audio generation technology. The model has "director-level" intelligence and can generate 5-second, 8-second, or 10-second multi-shot narrative units with one click, automatically completing natural transitions between long shots, medium shots, and close-ups. The innovation lies in synchronously generating human voice, ambient sound, and emotional dubbing, making character lip movements and actions perfectly match the audio. Source: AIsphere

December 1 · Monday

DeepSeek V3.2 official version: strengthens Agent capabilities, incorporates thinking and reasoning

DeepSeek officially releases the DeepSeek-V3.2 version, including the standard version and the Speciale long-thinking enhanced version. The standard version balances reasoning capability and output efficiency, reaching the GPT-5 level; the Speciale version wins gold medals in international competitions such as IMO and ICPC, with reasoning capability comparable to Gemini-3.0-Pro. The upgrade realizes the integration of thinking and tool calling for the first time, reaching the highest level among open-source models in Agent evaluations. Source: DeepSeek

ByteDance's video model Vidi2 surpasses Gemini 3 Pro! Comprehension capability off the charts

ByteDance releases the new-generation video understanding model Vidi2, whose performance on core tasks such as spatiotemporal localization surpasses GPT-5 and Gemini 3 Pro. The model can accurately understand hours-long video content and directly generate a complete JSON editing plan including editing time points, subtitles, music, and other details, realizing AI-automated editing from raw footage to finished film. Source: Quantum Bit

StepFun open-sources the 4B Agent model GELab-Zero, running on all Android devices

StepFun open-sources GELab-Zero, launching the 4B-parameter GUI Agent model GELab-Zero-4B-preview. The model achieves same-size SOTA results on multiple benchmarks and can smoothly execute complex and ambiguous instructions. The project also open-sources the complete inference engineering infrastructure, supporting one-click deployment and multi-device task distribution, solving the engineering challenges of scaling mobile Agent deployment. Source: Quantum Bit

November 28 · Friday

DeepSeek makes a strong comeback, open-sourcing the IMO gold-medal-level math model DeepSeek-Math-V2

DeepSeek open-sources the new-generation math model DeepSeek-Math-V2, achieving IMO gold-medal-level reasoning capability. The model breakthrough introduces a self-verification mechanism, using a "generator-verifier-meta-verifier" triple architecture to give AI the ability to check and correct its own reasoning process. It scores 83.3% on the IMO 2025 test and nearly full marks on the Putnam 2024 test. Source: Machine Heart

Kunlun Wanwei's AI music platform Mureka officially releases the V7.6 and O2 dual models

Kunlun Wanwei's AI music platform Mureka officially releases the V7.6 and O2 dual models, pushing AI music creation into a new stage of "professional quality × real-time experience." The new models are based on the MusiCoT fine-grained music modeling system, achieving significant improvements in arrangement structure, semantic mapping, and sound field quality, and supporting generation in 10 languages and custom singer functionality. Source: Kunlun Wanwei Group

Hunyuan 3D Studio upgraded, can directly output artist-grade 3D assets

Tencent Hunyuan 3D Studio is upgraded to version 1.1, integrating the all-new art-grade 3D generation model PolyGen 1.5 and achieving a breakthrough in end-to-end native quad mesh generation technology. The model pioneers adaptive hybrid topology capability, can directly generate professional-grade 3D assets with regular topology, and surpasses existing SOTA methods in key metrics such as topology quality and structural integrity. Source: Tencent Hunyuan