Top AI News Weekly
Hollywood Draws a Line + Agents Hit a Ceiling. The Motion Picture Association and ByteDance signed a memorandum on 17 August putting guardrails on Seedance and Seedream and their outputs across TikTok, CapCut and Dreamina - and it is worth reading precisely, because it is a truce, not a settlement: no studio is being paid for what those models already learned. On capability, NVIDIA's AVO cleared all 183 levels of ARC-AGI-3 for a perfect 100.00 RHAE using 12% fewer actions than VISTA on Claude Opus 5, and made the point that system design rather than raw model strength carried it. Ornith-1.5 went further in the same direction, training a model to propose its own tasks and build its own scaffolds. Video kept compressing: AlayaWorld shipped interactive world generation with mid-stream prompt switching, and Qwen-Video-Edit showed you can edit video by repurposing an image model, no video transformer required. Quietly the most useful release for small teams was Comfy MCP going local, which lets an agent read your GPU, fetch the right weights and run ComfyUI on your own machine.
24 launches and research drops that matter for enterprise AI builders—curated, tagged, and ready for your next roadmap sync.
New drops
24
Unique sources
24
Key themes
Video · 3D · Audio
LLMs, reasoning models, MoE architectures, and language understanding systems.
A model trained to improve itself: it proposes new tasks, builds the scaffold each task needs, then generates solution rollouts, with all three stages optimised jointly under GRPO. Ships as 397B MoE, 35B MoE with 3B active, and a 9B dense with a quantised mobile build, on Qwen3.5 and Gemma 4 bases. The 397B posts 86.1 on Terminal-Bench 2.1 and 92.8 on GPQA Diamond.
View release ↗Image generation, editing, restoration, OCR, and visual understanding.
Built on the NEO-unify architecture, which drops both the visual encoder and the VAE so understanding and generation share one representation space. The 8B-MoT label means roughly 8B understanding parameters alongside 8B generation parameters. Does native 4K image generation with stronger Chinese and English text rendering, placed against Qwen-Image 2.0 Pro and Seedream 4.5.
View release ↗Image understanding arrives on deepseek-v4-flash-vision-exp, taking JPEG, PNG, GIF and WebP by base64, HTTPS URL or Files API. Limits are unusually generous: up to 600 images per request, 32MiB per image (64MiB via Files), 48MiB per request body, and 8,192 pixels per side falling to 4,096 once you pass 15 images. Each image costs up to 384 tokens, and images are accepted in user messages only.
View release ↗A running index of bit-depth enhancement research - showing low bit-depth content on higher bit-depth displays without false contouring or colour banding. Collects 40+ papers from 1998 to 2023 with code, split across image-domain CNNs (BE-CNN, BE-CALF, RMFNet) and video-domain spatiotemporal networks (SSCNN, TANet, 3D-BitNet), plus the Sintel, MIT-Adobe 5K and Kodak datasets.
View repo ↗Video generation, editing, inpainting, and real-time rendering.
Interactive autoregressive world model generating long-horizon video with real-time camera control and prompt switching mid-generation. Built on LTX-2.3 through a four-stage pipeline, holding spatial memory by reprojecting a 3D cache and temporal continuity through frame-history embeddings, then distilled few-step with DMD. v1.1 weights landed 17 August, a browser demo on the 20th. 833 stars.
View repo ↗Splits generation in two: Qwen2.5-VL-7B plans the semantic change from the instruction, then Wan2.2-T2V-A14B renders it - roughly 21B across the pair. Covers text-to-image, image-to-image, text-to-video, video-to-video and reference-guided variants. Scores 8.02 on EditVerse, 63.83 on OpenS2V and 84.46 on VBench. Apache-2.0, weights out.
View model ↗Instruction-based video editing without any video-pretrained transformer: it repurposes Qwen-Image-Edit to operate directly on Wan 2.1 video-VAE latents, bridged by two small projection layers. Latent frames are tiled into a virtual image so the pre-trained positional encoding still applies, with grid RoPE across frames. Fine-tuned by LoRA on Ditto-1M triplets. Code and weights released.
View project ↗A Tencent-hosted creation platform pitched simply as turning your imagination into a film. The public page carries the positioning and little else - no model names, capability list, pricing or launch date is disclosed yet - so treat it as a signal that Tencent is packaging its Hunyuan generation stack into a consumer film tool rather than as a specified product release.
View release ↗Motion design tool aimed at Lottie, the vector animation format that ships to iOS, Android and web. Imports from Figma and After Effects, edits animations directly, converts SVG to Lottie and Lottie out to GIF, MP4 or WebM, and adds a Magic Animator for AI-assisted motion. Real-time collaboration with commenting, plus hosting and file optimisation. Claims 500,000+ designers.
View release ↗TTS, STT, voice cloning, music generation, and audio processing.
Moved from Week 33. Bilibili zero-shot TTS clones a voice from one reference clip with emotion held separate from speaker timbre, across Chinese, English, Japanese, Spanish and Arabic. Roughly 0.8B on a GPT backbone, 22.05kHz out, rate adjustable 0.5-2.0x, pronunciation steerable via Pinyin, CMU phonemes or Kana. Runs in about 6GB of VRAM.
View model ↗Positioned as the smallest zero-shot TTS worth running: about 170M for the generative model plus 120M for the codec decoder. Chinese and English are primary, with German, Spanish, French, Italian, Japanese and Korean experimental. CV3 error rates are 3.619 WER Chinese and 3.307 English, with Japanese lagging at 12.322. Community licence is free below $2M revenue.
View model ↗Text-to-song generation in beta: describe the music and it writes a full composition, with a toggle between instrumental and Smart Lyrics and up to 5,000 characters of direction per track. Style presets span jazz, pop, R&B, rock, acoustic folk-pop and Mandopop. Pricing is not published yet, which makes it hard to place against MiniMax Music 3 or Suno on cost.
View release ↗3D generation, reconstruction, PBR materials, depth estimation, and spatial computing.
Builds a photorealistic 4D human from one casual monocular video - no rig, no calibration, no tripod. Reference Context Packing compresses a linearly growing reference context into a fixed budget, while Target Context Routing passes structure between otherwise disjoint generation groups. Generated views feed 4D Gaussian Splatting via FreeTimeGS. SIGGRAPH Asia 2026; code released.
View project ↗Reconstructs a 3D Gaussian head from multi-view capture in one forward pass, no test-time optimisation - an encoder compresses the views into a compact latent decoding to UV-parameterised Gaussians on a neutral head template. Trained on an internal set of 10,000+ subjects and evaluated on Ava-256 at 16, 6 and 4 views, beating Avat3r. ECCV 2026; no code released.
View project ↗Reconstructs long RGB video into globally consistent cameras and point clouds without losing local accuracy. A geometric prior model reads overlapping chunks to predict depth, confidence and camera parameters; test-time adaptation then does global Sim(3) alignment and hierarchical refinement. On Oxford Spires it reaches 0.575 degrees rotation error and 4.372m trajectory error.
View project ↗Markerless motion capture that turns multi-camera video into 3D animation data with no suits or markers, claiming quality comparable to optical systems at lower cost and faster shoot times. Founded 2019 on a patented motion engine, with scalable capture volumes. Lists Ubisoft, SEGA, Warner Music Group, Paramount, Nike, Sony, Meta, Disney and Unreal Engine among users. Pricing is quote-only.
View release ↗An India-registered browser-based design tool that puts an AI engine behind CAD-style creation, requiring no local install. The public page is currently positioning only - no capability list, model details, pricing or launch date is published - so this is worth watching as an Indian entrant in AI-assisted design rather than evaluating on specifications that do not yet exist publicly.
View release ↗Agent frameworks, orchestration, coding agents, desktop agents, and embodied AI.
Agentic Variation Operators cleared all 183 levels across 25 game environments for a perfect 100.00 RHAE on the ARC-AGI-3 public set, using about 12% fewer environment actions than VISTA on Claude Opus 5 (6,624 against 7,542). A supervisor watches the search trajectory while the agent plans, edits and evaluates against persistent memory. Published 21 August, arXiv 2603.24517.
View release ↗Embodied foundation model announced 19 August taking a 30-second video context plus sensor, language and proprioceptive input, emitting 100Hz action trajectories. Trained continuously over eight months on real captures from homes, warehouses and factories, with no simulation data in pre-training. One-shot success averages 59% across ten tasks; ten gradient steps lifts it to 83%.
View release ↗Desktop app bundling chat, image and video generation and a coding agent over 12 local backends (Ollama, LM Studio, vLLM, KoboldCpp and others), with ComfyUI built in for media. Adds 28 agent tools plus MCP, local-embedding RAG, voice in and out, A/B model comparison and phone access over a web app. No telemetry, fully offline after download, and it supports abliterated weights. AGPL-3.0, v2.6.6.
View repo ↗SDKs, inference engines, data pipelines, monitoring, and AI infrastructure.
Released 18 August and fully open source: an MCP server that lets Claude, Cursor or Codex drive a ComfyUI install on your own machine. The agent reads your GPU to work out which models will fit, pulls the weights, enumerates every installed node including custom ones, then builds and runs the workflow locally. Sits alongside the existing cloud MCP rather than replacing it.
View release ↗Harvey open-sourced three legal datasets: Legal Agent Bench with 1,200+ tasks across 24 practice areas graded against 75,000+ expert-written rubric criteria, LAB Contracts for negotiation, and LAB Diligence, an RL environment over data rooms reaching 80 million tokens. Research lead Gabe Pereyra frames it against the ImageNet precedent - open sets are how a field finds its own flaws.
View release ↗Research papers, alignment, red-teaming, interpretability, and benchmarks.
A memorandum signed 17 August putting global guardrails on how ByteDance models use film and TV property - covering Seedance video, Seedream image, and outputs across TikTok, CapCut and Dreamina. It follows February, when the MPA demanded restrictions after Seedance produced realistic AI actors. Notably not a licensing deal: no studio is paid for what the models already learned.
View release ↗A 350-entry index of agent-harness engineering, 90% of it actual GitHub projects rather than blog posts, sorted into nine buckets: harness architecture, context and working-state, execution sandboxes, tool protocols, evaluation harnesses, observability, guardrails, reference implementations (85 alone) and ecosystem maps. Last verified 13 August 2026. 1.6k stars.
View repo ↗