Skip to content

Content Formats

Model Releases Latest

Video/Image/Audio Generation Model Releases, Open-Sourcing, and Iterations: First-hand coverage of flagship updates, open APIs, and performance and pricing changes.

123 picksLast 30 days 45 itemsCollected 367 items

updated

Pick archive · page 4

Oct 15Thu61–80
  1. 量子位71

    Kunlun Wanwei's Skywork AI releases Matrix-Game 4.0, running in real time on a single RTX 5090 and open-sourced

    Skywork AI, under Kunlun Wanwei, releases Matrix-Game 4.0, which it calls the world's first game world model that can run in real time on a single consumer-grade graphics card, and has open-sourced the model and code. The model leaves game rules to the game engine to maintain and visuals to the video model to generate; when a player is hit, shield durability is deducted before health, and the rules can be rewritten by a Coding Agent using natural language.

    Why it matters: Running a game world model in real-time on a single RTX 5090, with rules handed to the game engine and visuals to the video model, providing a reproducible engineering pathway for world model deployment.

  2. HeyGen Updates78

    HeyGen releases the HeyGen Voice speech model

    HeyGen releases a new speech model, HeyGen Voice, which it calls the first speech model built for video adaptation and which ranks first on the Artificial Analysis speech leaderboard; it is already included in every HeyGen video at no extra charge.

    Why it matters: HeyGen launches its first speech model for video adaptation, allowing creators to assess the room for replacing dubbing and multilingual lip-sync workflows.

Oct 14Wed
  1. OpenAI News60

    Higgsfield uses GPT-6 Astra to extend AI video from 1 minute to over 10 minutes

    Higgsfield extends AI video generation length from 1 minute to over 10 minutes, powered by GPT-6 Astra. At the same time, it uses GPT-Image-2.5 to improve image precision and GPT-Live to provide a path toward voice-driven filmmaking.

    Why it matters: Higgsfield uses GPT-6 Astra to push AI video length from 1 minute to over 10 minutes, which can be used to assess its usability for feature-length creation.

  2. Fish Speech 官方发布72

    Fish Audio releases the open-source speech generation model Drama 3 pre-release version

    Fish Audio releases the open-source speech generation model Drama 3 pre-release version, which can convert text into performed speech. The model uses a 35B-A3B hybrid linear attention MoE backbone (initialized from Qwen3.5) and a new codebook-MTP audio decoder; pretraining covers about 80 languages and dialects and hundreds of billions of tokens, and it supports multi-reference voice cloning, natural language instructions, multi-round revisions, multi-character scenes, speech infill editing, and long-audio and streaming synthesis.

    Why it matters: Open-source speech models put multi-character scenes, multi-round script revisions, and voice editing into the same model, allowing dubbing workflows to assess replacement potential accordingly.

Oct 13Tue
  1. Black Forest Labs · FLUX 模型78

    Black Forest Labs releases FLUX 3 Image Klein open-weight model

    Black Forest Labs Releases FLUX 3 Image Klein, a 5.6B Open-Weight Model That Simultaneously Supports Text-to-Image, Single-Reference Image Editing, and Multi-Reference Image Combination Generation, with Qwen3-VL-4B-Instruct as the text encoder.

    Why it matters: The 5.6B open-weight model unifies text-to-image, single-image editing, and multi-image combination into a single model, and provides quantized versions with VRAM usage details, making it easier for self-deployers to assess implementation costs.

  2. Black Forest Labs · FLUX 模型76

    Black Forest Labs Open-Sources FLUX 3 Image Klein Base, 5.6B Weights Support Text-to-Image and Multi-Image Editing

    Black Forest Labs has released FLUX 3 Image Klein Base on Hugging Face, an open-weight model with 5.6B parameters. Text-to-image, single-reference image editing, and multi-reference image composition share the same model, with the text encoder being Qwen3-VL-4B-Instruct.

    Why it matters: The 5.6B open weights cover text-to-image, single-image editing, and multi-image composition simultaneously, and provide three tiers of VRAM usage—BF16, FP8, and NVFP4—making it easy to evaluate self-hosting costs.

  3. ComfyUI Wiki · 工作流与开源创作分享78

    Fish Audio releases Drama 3 speech synthesis model, supporting 16 characters and natural-language performance instructions

    Fish Audio released its next-generation speech synthesis model Drama 3 on October 13, 2026, an open-weight 35B MoE model activating 3B per token.

    Why it matters: The 35B open-source speech model supports same-track generation of 16 characters and natural-language performance instructions; its pronunciation robustness data can be used to assess dubbing and audiobook workflows.

Oct 12Mon
  1. 腾讯 ARC · 创作模型60

    Tencent ARC Open-Sources Multi-Person Video Generation Model Multics

    Tencent ARC released Multics on Hugging Face, achieving scalable multi-person video generation through layered communication. The repository includes the complete 4-player Multics DiT checkpoint, configuration, Wan2.2 VAE, UMT5 text encoder, and tokenizer; the same checkpoint supports 6-, 8-, and 10-person generation without retraining. The weights are distributed under the Apache 2.0 license.

    Why it matters: Tencent ARC open-sources the multi-person video generation model Multics, where the same weights can scale to 10 people, letting creators evaluate generation options for multi-person scenes.

Oct 9Fri
  1. SuSu_酥酥66

    Alibaba Tongyi open-sources Qwen-Image-2.1-Turbo, producing 2K images in 8 steps

    Alibaba's Tongyi open-sources Qwen-Image-2.1-Turbo, an accelerated checkpoint based on Qwen-Image-2.1 that uses the same 7B visual generation architecture; text-to-image and natural-language editing require only 8 denoising steps, and the official release says it can still generate 2K images and supports continuous editing such as adding accessories and changing scenes.

    Why it matters: Qwen-Image-2.1-Turbo achieves 2K text-to-image and natural-language editing with 8 denoising steps, and offers two onboarding paths: Diffusers and API.

  2. 通义千问 · 图像与音频模型62

    Alibaba releases Qwen-Image-2.1-Turbo accelerated image generation model

    Qwen released Qwen-Image-2.1-Turbo, an accelerated checkpoint of Qwen-Image-2.1 for text-to-image and image editing, requiring only 8 denoising steps.

    Why it matters: 8-step sampling and KV cache reuse of text and reference image context let creators assess the speedup potential of their image generation and editing workflows.

Oct 8Thu
  1. 机器之心 · 公开公众号 RSS73

    Shengshu Technology Releases Vidu Q4 Preview, Starting at 0.09 Yuan/Second

    Shengshu Technology has launched Vidu Q4 Preview, the first preview version of Vidu Q4 for creators, focusing on enhancing character performance, camera language, and complex visual effects. The launch price starts at 0.09 yuan per second, allowing up to 5 times more content to be generated with the same budget.

    Why it matters: Vidu Q4 Preview combines a launch price of 0.09 yuan per second with 15 reference images and 2K/4K output, providing a basis for judging the cost range of AI video mass production.

Oct 7Wed
  1. Oimi · 中文图像视频教程67

    Google releases Nano Banana 2.1: half the price, panoramic aspect ratios, and 14-image fusion

    Google officially released the image generation and editing model Nano Banana 2.1 (model ID: gemini-nano-banana-2.1) in the Gemini API, at about $0.034 per 1K image, roughly half that of Nano Banana 2.

    Why it matters: Google's image model 2.1 cuts the per-image cost by about half and adds panoramic aspect ratios and 14-image fusion, which can be used to evaluate whether to migrate an existing image-generation pipeline.

  2. The Decoder70

    Google releases image model Nano Banana 2.1, priced at roughly half

    Google released the image generation and editing model Nano Banana 2.1, replacing Nano Banana 2 released in February 2026, with the company claiming improvements across the board. The price of 1K images dropped from 6.70 cents to 3.36 cents, and 4K from 15.10 to 7.56 cents, roughly half of the previous generation.

    Why it matters: Google's image model Nano Banana 2.1 comes with multiple benchmark comparisons while cutting prices by about half, which can be used to judge whether to replace your existing image generation pipeline.

  3. SuSu_酥酥78

    Google Releases Nano Banana 2.1 Image Generation and Editing Model

    Google has officially released the Nano Banana 2.1 image generation and editing model, which the company says comprehensively surpasses the previous generation in visual design, precise mask editing, and subject consistency, with more natural and realistic images. This version has rolled out across Gemini, Search AI Mode, AI Studio, Flow, Stitch, Google Ads, and enterprise platforms.

    Why it matters: Google's image generation model gets an upgrade, calling out changes in three capabilities—visual design, mask editing, and subject consistency—and has rolled out across multiple product lines.

Oct 6Tue
  1. FastVideo 视频模型 · 模型发布62

    FastVideo Releases FastH3 Trim Lightweight Model and ComfyUI Repackaged Files

    FastVideo released the experimental lightweight model FastH3 Trim, reducing transformer blocks from 50 to 42, using rank-16 timestep conditioning, 8-step sampling, with a 4.2x smaller size than the base H3 and faster than FastH3 V2.

    Why it matters: FastH3 Trim offers a lighter video generation solution with 42 layers, 8-step sampling, and 4.2x size compression, along with ComfyUI files and parameters for various precisions.

Oct 5Mon
  1. GitHub:AI视频开源项目发现82

    Kandinsky 6.0 open-sourced: synchronized audio-video generation models, in Lite and Pro tiers

    kandinskylab released the Kandinsky 6.0 Video series of diffusion models, including a 3B-parameter Lite and a 29B-parameter Pro, capable of generating 5-second videos with 44 kHz synchronized audio (including lip-sync), supporting text-to-audio-video and image-to-audio-video, and paired with a super-resolution model that upscales output to Full HD (1920×1080).

    Why it matters: Kandinsky 6.0 has open-sourced video generation models with synchronized audio and lip-sync, and provides measured runtimes for various GPU tiers, making it easy for creators to assess local deployment costs.

Oct 3Sat
  1. FastVideo 视频模型 · 模型发布62

    FastVideo releases the FastH3 Trim 8-step audio-video generation model

    FastVideo released the experimental model FastH3 Trim, which removes 8 of the least impactful blocks from FastH3 V2's 50 MiniMax H3 transformer blocks and replaces each block's AdaLN timestep projection with a shared rank-16 basis, trained with 8-step DMD2 (checkpoint 300), enabling 8-step text-to-video generation with synchronized audio.

    Why it matters: FastH3 Trim uses pruning plus shared AdaLN to bring 8-step audio-video generation to consumer-grade hardware, and can be compared against V2 to weigh image quality against size.

  2. IndexTTS · 语音模型60

    IndexTeam Releases Index-Translate-35B-A3B-preview Multilingual Translation Model

    IndexTeam has released Index-Translate-35B-A3B-preview, a multilingual text translation model based on Qwen3.5 with 35B total parameters and 3B active parameters in an MoE architecture. It supports 150 languages and natively supports end-to-end image translation without requiring external OCR.

    Why it matters: The Index-Translate family adds a 35B-A3B preview version supporting 150 languages and direct image translation, allowing evaluation of its low-resource language performance against report data.

Oct 2Fri
  1. 可灵 Kling72

    Kling 4.0 enters closed beta, official release in October

    Kling 4.0 has entered closed beta, with the official release scheduled for October this year. The new version upgrades audio and visuals, supporting stable dynamic motion, high-quality stereo, more precise lip-syncing, up to 4K resolution and 10-bit HDR output, and can generate videos up to 30 seconds long.

    Why it matters: Kling 4.0 has entered beta testing, allowing creators to understand its changes in dynamic stability, long-take capability, and consistency.

  2. 可灵官方博客62

    Kling 4.0 enters closed beta: creators test 30-second long takes and multi-shot consistency

    Kling 4.0 has entered closed beta, with an official release planned for October this year; upgrades include stable dynamic motion, high-quality stereo audio, more accurate lip sync, up to 4K resolution and 10-bit HDR output.

    Why it matters: Kling 4.0 enters closed beta; creators test 30-second long takes, motion stability and multi-shot consistency, which can help judge whether a commercial production pipeline is worth switching to.