New 3D editors, open medical AI, AI symphony, Qwen 3.8, Wan Animate 2: AI NEWS

SYSymphonyGenUnverified

Generates full orchestral music by first producing a harmony skeleton then expanding it into a complete arrangement. Accepts custom or extracted harmony skeletons; open-source models under 5 MB.

MMMAC (Multi-Agent CAD)Available in EU

Multi-agent system that turns text prompts into printable CAD files. Claims 116x fewer tokens and 13x lower cost than CAD Skills with a higher pass rate; model agnostic and runs locally.

WAWan Animate 2AlibabaUnverified

Animates a character photo using a reference video, with hands, faces, multiple characters, non-human characters and camera angle control. A Light version supports real-time streaming; ComfyUI support and an int8 version exist.

VOVocalRenderUnverified

Singing voice generator that takes lyrics and MIDI melody and outputs expressive vocals, using an autoregressive plus diffusion design. Two open models under 10 GB, with training code; trained only on Chinese so far.

HBHunyuan3D Buffalo 1.0TencentUnverified

Unified model that generates, understands, edits via text and segments 3D objects into parts. Code is listed as coming soon.

LELeapTalkUnverified

Real-time talking avatar generator from an image and speech audio. Motion is rigid but it is extremely fast, claimed up to 200 fps on an H200; released openly.

Q3Qwen 3.8 MaxAlibabaAvailable in EU

2.4T-parameter model, the first Max-class Qwen to be open-sourced with weights promised the following week. Shown running long autonomous tasks over days and ranking just below Kimi K3 on Artificial Analysis; available via API.

W2WeatherNext 2Google DeepMindUnverified

Single model predicting cyclone track, intensity and wind structure up to 15 days ahead with 1000-scenario ensembles. Gives 24+ hours more lead time than leading systems; open-sourced with a Mini version for Colab.

GAGPT-6 AstraOpenAIUnverified

An internal version of OpenAI's upcoming model, codenamed Astra, solved or advanced 10 open math problems at about $2,000 of API-rate compute. Reasoning walkthroughs were published.

CLClinFusionAlibaba DAMO AcademyAvailable in EU

Open medical multimodal model handling 2D and 3D imaging plus text for question answering and clinical reports. Comes in 32B and 8B versions and beats models like GPT-5.2 on benchmarks.

PAPersona AI Gen 1 humanoid weldingPersona AIUnverified

Demo of the Gen 1 humanoid performing a welding task via VR teleoperation, pointing to use in dangerous industrial settings.

USUBTECH swarm intelligence (Cruiser Y1)UBTECHUnverified

Preview of a swarm intelligence system coordinating multiple wheeled industrial humanoid robots in a warehouse without overlap.

XRXiaomi Robotics 1XiaomiAvailable in EU

Robot foundation model following natural-language instructions for tasks like packing a suitcase. Trained on ~100,000 hours of handheld-gripper human video plus ~10,000 hours of robot data; released openly.

BBBig Bang v1Available in EU

Experimental language model based on Qwen 3.6 35B, post-trained on data from generator, critic and metacritic agents. Scores much higher than the base on benchmarks such as Frontier Science; released openly.

Closed-source Meta model focused on coding and agentic workflows with a 1M token context and multimodal input. Self-reported benchmarks look strong but independent leaderboards place it low, though it is cheap; API only.

Meta's coding agent harness, analogous to Codex or Kimi Code, recommended as the best way to run Muse Spark.

LHLong Horizon HarnessUnverified

Harness using manager, executor and auditor roles to keep long-running agent tasks on track. Works with Claude Code, Codex CLI and others, raising scores on Terminal Bench and other benchmarks at some extra token cost.