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

This week’s AI roundup showcases an extraordinary wave of breakthroughs spanning open-source models, creative tools, and research—from Alibaba’s powerful Qwen models to Open AI’s reported math breakthroughs and Google’s open-source cyclone predictor. New creative tools let users generate full orchestral music with control over harmony, turn text prompts into printable 3D CAD models, animate still characters from reference videos with real-time options, and produce strikingly realistic singing voices. Tencent’s 3D generator, a lightning-fast talking-avatar system, and a 30-second narrative video model further push creative boundaries, while open-source frontier models like Kimi K3 and Qwen 3.8 Max now rival or exceed closed competitors. Agentic autonomy reached new heights as Qwen 3.8 Max ran self-improving loops for 16 days, reproduced research, and designed chip layouts without human help. Overall, the week underscores how rapidly open-source AI is closing the gap with proprietary systems across both capability and accessibility.

• ▶ 0:00 The host calls the week in AI "absolutely insane," previewing a dense roundup across research, open-source, creative tools, and robotics.
• ▶ 0:07 A new open-source AI for medical research and reporting is announced, alongside Alibaba's latest Qwen model, described as "an absolute beast."
• ▶ 0:28 OpenAI's internal model reportedly solved massive math breakthroughs, and Google open-sourced a state-of-the-art AI for predicting cyclones and natural disasters.
• ▶ 0:50 Symphony Gen is a new AI tool that creates full orchestral music while giving users control over the underlying harmony, rather than a fixed autonomous composition.
• ▶ 1:47 It works by first generating a harmony skeleton, then expanding it into a complete orchestral arrangement—like sketching chords before filling in each instrument's notes.
• ▶ 2:25 Users can input their own harmony skeleton or extract one from an existing piece to generate a new inspired piece; the released models are very lightweight (under 5 MB) and run locally on most devices.
• ▶ 3:30 MAC (Multi-Agent CAD) is an AI tool that turns a simple text prompt into a printable 3D model in CAD format.
• ▶ 4:06 The system is highly efficient and cost-effective, completing tasks at roughly 10 times lower cost; compared to the CAD skills model, it uses 116 times fewer tokens, costs 13 times less, and achieves a much higher pass rate.
• ▶ 4:32 Setup is super simple and model-agnostic (e.g., Qwen), with all instructions provided online for downloading and running MAC locally on your own computer.
• ▶ 4:54 Alibaba has released One Animate 2, a new animation system.
• ▶ 4:59 It takes a photo of any character and a reference video, then animates the character to follow the reference motion.
• ▶ 5:06 One Animate 2 transfers motion from a reference video to a still image, including hands, fingers, and non-human characters like a teddy bear.
• ▶ 5:19 It also transfers facial expressions and supports animating multiple characters, even from a single-person reference video to a multi-character photo.
• ▶ 6:04 A smaller variant, Wan Anime 2 Light, enables real-time streaming with under-one-second latency for live use.
• ▶ 6:17 One Animate 2 is a significant upgrade over previously covered tools like One Animate and Dream Actor, offering more detailed, consistent, and natural results.
• ▶ 6:27 The host compares One Animate 2 to Scale 2, noting the quality of both tools is “very similar,” even though the developers did not directly compare them.
• ▶ 6:35 One Animate 2 is already released, with two versions: a full ~33 GB version for high-end GPUs and an INT8 version half the size for mid-tier GPUs, plus Comfy UI support.
• ▶ 6:55 Vocal Render is introduced as a new AI tool that generates singing voices for songs, with output described as incredibly realistic and expressive.
• ▶ 7:06 The model takes lyrics and a melody (as MIDI notes) as inputs, then outputs an expressive sung vocal.
• ▶ 7:17 Two variants are released — Vocal Render and Vocal Render Pro — with the Pro version sounding "a bit better" than the standard version.
• ▶ 7:21 Vocal Render's demo impresses immediately, with a gasp at ▶ 7:23, and the pro version at ▶ 7:30 draws another strong reaction at ▶ 7:54.
• ▶ 7:39 Direct comparison against Vivo 2 and Soul X leads to the host's verdict at ▶ 7:45 that Vocal Render is "a lot better."
• ▶ 8:01 A key technical flaw is exposed: Vivo 2 failed to follow the specified pitch, while Vocal Render maintained pitch accuracy.
• ▶ 8:03 Vocal Render reads lyrics and musical notes together to predict performance flow, automatically deciding final timing and audio length—critical when one syllable stretches across multiple notes.
• ▶ 8:20 It uses a two-stage architecture: an autoregressive component first builds a broad sketch of the singing style and timing.
• ▶ 8:27 A diffusion model then fills in finer details, including pitch, vocal tone, articulation, and local audio texture, enabling expressive, timing-accurate vocals.
• ▶ 8:33 Vocal Render has been publicly released, with a "View Repository" button providing all instructions for downloading and running the model locally.
• ▶ 8:44 The training code is also released; the pre-trained model is Chinese-only, but users can train their own checkpoint in any language.
• ▶ 9:01 Two model variants (pro and normal) are both under 10 GB, fitting on most consumer GPUs, and at ▶ 9:08 the speaker recommends Vocal Render as one of the best models for realistic, expressive AI singing voices.
• ▶ 9:22 Tencent released Hunyan 3D Buffalo, a unified 3D model that can generate, edit, and separate 3D objects from text prompts, with code coming soon.
• ▶ 10:44 Leap Talk generates real-time talking avatars from any reference image and speech audio, and is thousands of times faster than other generators (up to 200 fps on H200).
• ▶ 12:10 Higsfield's Seedance 2.5 can produce up to 30 seconds of video in a single pass with narrative and audio, while accepting up to 50 references for strong consistency.
• ▶ 13:31 Open-source frontier models surged: Moonshot AI's Kimmy K3 (2.8T params) matched or exceeded closed models like GPT 5.6, and Alibaba's Quen 3.8 Max (2.44T) marked the first open-sourced "max" class model with weights coming next week.
• ▶ 14:40 Agentic autonomy reached new heights: Quen 3.8 Max ran an autonomous self-improving harness for 16 days (265 commits, 127 PRs), reproduced and beat a research paper by 2.7 points, and designed a chip layout 12x smaller than baseline without human help.
• ▶ 16:04 On the Artificial Analysis leaderboard, Quen 3.8 Max sits just one point below Kimmy K3 despite being ~400B params smaller, nearing GPT 5.6 and Claude Fable—though confidence intervals are missing, so top-model gaps may not be significant.
• ▶ 16:56 Google DeepMind's WeatherNext 2 combines storm track, intensity, and wind forecasting into one open-source model, generating 15-day forecasts and 1,000-scenario ensembles in under a minute on a single TPU.
• ▶ 19:24 OpenAI's rumored GPT-6 (Astra) reportedly solved 10 long-standing open math problems across fields like group theory and coding theory, with an estimated token cost of only $2,000.
• ▶ 21:19 Alibaba's Damo Academy released Clinfusion, an open-source medical AI model designed to take in X-rays, scans, and other medical imagery for holistic understanding.
• ▶ 23:01 Persona AI demonstrated teleoperated welding with its Gen 1 humanoid, highlighting remote expert control for high-risk industrial tasks.
• ▶ 23:52 UB Robotics showed swarm intelligence with multiple Cruiser Y1 robots working together in a warehouse, coordinated by an overarching system to avoid redundancy.
• ▶ 25:01 Xiaomi’s Robotics 1 model was trained on ~100,000 hours of handheld camera/gripper video plus ~10,000 hours of real robot data, and was open-sourced to support broader robotics development.
• ▶ 25:44 Big Bang is an experimental "self-evolving" model that uses generator, critic, and meta-critic AI agents to automatically create and filter harder training problems, starting from Qwen 3.6 35B.
• ▶ 26:29 On benchmarks, Big Bang shows major jumps over its base model—especially Frontier Science leaping from ~12 to ~46—though "self-evolving" is misleading and performance will eventually hit a ceiling.
• ▶ 27:28 Meta quietly released Muse Spark 1.2, a closed-source, API-only coding-focused model with a 1M token context, but its self-reported benchmarks are misleading and independent leaderboards rank it far below top models despite being very cheap.
• ▶ 29:24 Meta also released Muse Code, a dedicated coding agent that the presenter recommends using as the optimal harness for running Meta's models.
• ▶ 30:41 The Long Horizon Harness uses a three-role workflow — Manager, Executor, and Auditor — to break long tasks into verified steps and prevent agents from losing the goal or making false completion claims.
• ▶ 31:28 Benchmark results show major gains: Qwen 3.7 in Claude Code +28.9% on WeaveBench, tripled OS World 2 completion rate, and +7.5% on Terminal Bench.
• ▶ 32:13 The harness trades tokens for performance (more tokens for WeaveBench/OS World, fewer for Terminal Bench) and has already been released, with download instructions linked in the code repo and description.

• ▶ 0:00 Intro & Creative AI Tools (0:00 - 9:18) - Weekly intro, AI symphony music, 3D modeling, One Animate 2, and Vocal Render singing AI.
• ▶ 9:18 AI News, Talking Avatars & Sponsor Spotlights (9:18 - 13:31) - Hunyan 3D Buffalo, Leap Talk avatars, and sponsor showcases for Higsfield and Seance 2.5.
• ▶ 13:31 Frontier Model Releases & Benchmarks (13:31 - 16:56) - Kimmy K3 and Quen 3.8 Max releases, agent demos, leaderboard, pricing, and access.
• ▶ 16:56 Weather, AI Rumors & Medical AI (16:56 - 22:51) - WeatherNext 2 hurricane forecasts, OpenAI Astra/GPT-6 math rumors, and Clinfusion medical AI.
• ▶ 22:51 Robotics & Embodied AI (22:51 - 25:40) - Persona AI welding, UB Robotics swarm intelligence, and Xiaomi Robotics 1 foundation model.
• ▶ 25:40 Self-Evolving Models & Meta AI Releases (25:40 - 30:01) - Big Bang self-evolving model, Meta Muse Spark 1.2, and Muse Code coding agent usage.
• ▶ 30:01 Long Horizon Harness & Closing (30:01 - 33:37) - Long Horizon Harness workflow, benchmarks, release access, and weekly wrap-up call to action.