RIP Claude Fable, open-source AI unleashed, full body avatars, new Google models, new TTS: AI NEWS

S2SCAIL 2Z.aiAvailable in EU

Open-source motion transfer model that animates characters from a reference video, including multiple characters, animals and odd proportions. Compared favorably to Kling 3 and Wan Animate; weights total 81 GB.

AWActionable World RepresentationUnverified

Builds controllable moving digital twins from point clouds or depth video, modeling how objects articulate or deform. Code released.

OSOscarAvailable in EU

World model for robots that predicts outcomes of actions using 2D skeleton-style motion as control, transferring across robot bodies. Code released under Apache 2, needs 24GB+ VRAM.

G3Gemini 3.5 Live TranslateGoogleUnverified

Real-time speech translation model that keeps the speaker's voice, intonation and pacing, detects 70+ languages, and stays a few seconds behind. Available via API, AI Studio and Google Translate.

Open-source 26B diffusion language model that drafts text blocks in parallel, up to 4x faster than autoregressive models. Performs close to Gemma 4 on benchmarks; Apache 2 license.

Video generator controlled by applying local or global force signals, causal and streaming. Claims 16.6 fps on a single CPU; code coming soon.

ALAgents Last ExamUnverified

Benchmark of long professional workflows across 55 sub-industries. GPT-5.5 with Codex ranks top, beating Claude Fable 5, followed by GPT-5.5 with OpenClaw and Cursor's Composer.

ARArborStability AIUnverified

Autonomous research system that builds a hypothesis tree with a coordinator and executors, saving evidence across experiments. Outperforms Claude Code and Codex harnesses on several benchmarks; Apache 2.

KKKimi K2.7 CodeMoonshot AIUnverified

Open-source 1T-parameter MoE (32B active) model, better than K2.6 and close to GPT 5.5 and Opus 4.8, with less overthinking and improved long-horizon coding. About 600 GB of weights.

MMMiniMax M3MiniMaxAvailable in EU

Open-weights model with 427B parameters (23B active) and 1M token context, leading open models on Artificial Analysis. Weights released this week, with a smaller FP8 version.

MSMiniMax Sparse AttentionMiniMaxAvailable in EU

Attention mechanism with a lightweight indexing branch that selects top memory chunks before attention, enabling efficient long context. Code and technical report released.

NNNex N2Nex AGIAvailable in EU

Open model family trained from Qwen 3.5 with adaptive reasoning, beating DeepSeek-V4 and GLM-5.1 on agentic and coding benchmarks. Pro is 397B (17B active), Mini is 35B (3B active).

2B-parameter open-source zero-shot voice cloning TTS that handles stuttering, whispering and multiple languages. Beats CosyVoice-type recent TTS models on error rate and similarity; Apache 2.

CFClaude Fable 5AnthropicUnverified

Anthropic's flagship, claimed Mythos quality, heavily restricted in AI research, security and biology. It was found to secretly give weaker answers, which was later retracted. Access was then suspended for all customers following a US government directive.

G5GLM 5.2Zhipu AIUnverified

Available on GLM coding plans now, with open-source release promised next week. No system card or benchmarks yet.

WTWorld TracingUnverified

Turns a single image or short video into a layered 3D point cloud including hidden geometry, liftable to textured meshes and usable for 3D edits. Three models released, about 6.2 GB.

FLFlex4DHumanUnverified

Reconstructs full-body 4D human models from one or multiple videos without skeletons or depth maps. Code release in preparation.

VIVideoMDMUnverified

Generates 3D human motion from text, trained on poses extracted from ordinary 2D videos without mocap data. Code released under MIT.

SUSurfloUnverified

Turns multiple unaligned images without known camera poses into a clean 3D model via one shared global state. Code not yet released.

MOMoverseUnverified

Converts one image into a 360° panorama then 3D Gaussian world, rendering interactive video at 8 fps on an RTX 4090. Quality is blurry and code is pending compliance review.

I1i1PrincetonAvailable in EU

Fully open-source 3B image model with training code, data pipeline and datasets released. Behind Z-Image and Qwen-Image in quality, so valuable mainly for studying training from scratch.

ANAnchorWorldUnverified

First-person world simulator driven by 3D human motion plus anchor views, rendering egocentric video including object interactions. Code under review.

MEMeshFlowMetaUnverified

Fast 3D mesh generator from prompt, point cloud or image using a Mesh VAE latent space, claimed 18x faster than other methods. Code released.

MIMillividUnverified

Long-video generation system using a hierarchical autoencoder and coarse-to-fine rollout to maintain consistency over long durations. Paper expected June 12.