DeepSeek AI launched the official DeepSeek-V3.2 on December 1, 2025 — a frontier large language model series optimized for superior reasoning and agent performance, now available via API, web, app, and Hugging Face. Introducing DeepSeek Sparse Attention for long-context efficiency, a scalable RL framework rivaling GPT-5, and a groundbreaking "Thinking in Tool-Use" mode that fuses step-by-step deliberation with external tool calls, V3.2 and its high-compute V3.2-Speciale variant achieve gold-medal IMO/IOI scores while slashing inference costs. This release marks DeepSeek's boldest agentic leap, empowering developers with seamless hybrid workflows.

ByteDance unveiled Vidi2 on December 1, 2025 — a 12-billion-parameter multimodal large language model specialized for video understanding and creation, now open-sourced on GitHub. Crushing benchmarks like VUE-TR-V2 (temporal retrieval) and VUE-STG (spatio-temporal grounding) with scores of 53.19% Temporal IoU — double that of Gemini 3 Pro's 27.50% — Vidi2 processes hours-long footage for precise edits, story-aware cuts, and highlight extractions. Leveraging TikTok's 1B+ daily users for training, it outperforms GPT-5 on long-video QA while running on consumer hardware, marking ByteDance's boldest strike in the video AI arena.

StepFun AI launched GELab-Zero on November 28, 2025 — a fully open-source 4B-parameter multimodal GUI agent model designed for autonomous Android control, now available on Hugging Face under Apache 2.0. Featuring plug-and-play infrastructure for ADB handling, task recording, and local inference on consumer hardware (as low as 16GB VRAM), it achieves 75.86% success on AndroidWorld benchmarks for complex tasks like ride-hailing and shopping. This zero-dependency beast runs offline on any Android device, slashing latency while preserving privacy — a direct challenge to cloud-heavy agents like Claude's Computer Use, empowering devs to automate mobile workflows without vendor lock-in.

DeepSeek AI roared back on November 27, 2025, with DeepSeek-Math-V2 — a groundbreaking open-source math reasoning model built on DeepSeek-V3.2-Exp-Base, achieving gold-level scores on IMO 2025 and CMO 2024, plus a near-perfect 118/120 on Putnam 2024 via scaled test-time compute. Featuring a dual verifier-generator architecture for self-verifiable proofs, it outpaces Claude 4 and Gemini on IMO-ProofBench, emphasizing rigorous step-by-step logic over mere answers. Weights dropped on Hugging Face under Apache 2.0, democratizing Olympiad-grade AI for researchers and educators worldwide.

Tencent unveiled Hunyuan 3D Studio 1.1 on November 28, 2025 — a seismic upgrade integrating the cutting-edge PolyGen 1.5 model for end-to-end quad-mesh generation, delivering production-ready 3D assets in minutes. Pioneering native quad topology learning with seamless edge loops and hybrid surfacing, it slashes wiring complexity by 70% for soft/hard models, outpacing traditional tools like Blender in efficiency and fidelity. Free for core use via the web platform, with pro API access on Tencent Cloud, this launch catapults creators from concepts to game-ready props, signaling Tencent's dominance in AI-driven 3D pipelines.

Black Forest Labs unleashed FLUX.2 on November 25, 2025 — a groundbreaking 32-billion-parameter rectified flow transformer that's now open-sourced under Apache 2.0, shattering benchmarks in text-to-image generation, single/multi-reference editing, and high-res photorealism up to 4MP. With variants like FLUX.2 [dev] for devs and [pro] for production, it crushes Midjourney V7 and Stable Diffusion 3 on prompt adherence (ELO 1,038) and visual quality, while enabling no-finetune character/style fusion from up to 10 refs. Hugging Face downloads hit 500K+ in 48 hours, signaling a seismic shift toward efficient, community-forged visual AI.

LiblibAI, China's premier AI creation platform, rolled out "Base Algorithm F.2" on December 10, 2025 — an upgraded generative engine that unlocks multi-image referencing for seamless style fusion and advanced in-painting/out-painting editing tools. Integrated into the 2.0 ecosystem alongside powerhouses like Qwen Image, Kontext, and Seedream 4.0, F.2 slashes iteration times by 70% for complex scenes, supporting up to 14 reference images per prompt with pixel-precise control. Free for all users (300 daily credits), it's already fueling 50K+ generations in 48 hours, positioning LiblibAI as the go-to for e-comm visuals, game assets, and VFX — outpacing Midjourney V7 in fidelity and flexibility.

Tencent launched the international version of its Hunyuan 3D Creation Engine on November 25, 2025 — a browser-based powerhouse at 3d.hunyuanglobal.com that generates high-fidelity 3D models from text, images, or sketches in under 5 minutes. Supporting OBJ/GLB exports, low-poly topology, and seamless integrations with Unity/Unreal/Blender, it offers 20 free daily generations for individuals and API access via Tencent Cloud for enterprises. Already powering 150+ Chinese firms like Unity China and Bambu Lab, this debut marks Hunyuan's overseas breakout, slashing 3D workflows by 90% and challenging Luma AI's throne in gaming, e-comm, and VFX.

On November 25, 2025, ByteDance's TRAE AI coding platform rolled out its flagship SOLO mode to the Chinese market — fully free and open to all users, no waitlists required after initial phased access. This "Responsive Coding Agent" fuses context engineering with multi-agent orchestration, handling everything from natural-language prompts to deployment-ready apps in parallel workflows, with features like DiffView for code diffs, auto-compression for mega-contexts, and seamless switches between IDE and SOLO modes. Since TRAE's China launch in March, it's amassed 1M+ MAU and 6B+ lines of adopted code; SOLO's debut promises to supercharge indie devs and startups, outpacing Cursor and Qwen IDE in automation depth.

Anthropic launched Claude Opus 4.5 on November 24, 2025 — its frontier flagship that's now the undisputed king of coding, agentic workflows, and computer use. Blending enhanced memory for endless chats, a tunable "effort" slider for 21% smarter reasoning at 66% lower costs, and integrations like Claude for Chrome/Excel, it smashes benchmarks: 80.9% on SWE-Bench Verified (first over 80%), 15% leap on Terminal Bench, and SOTA on GPQA Diamond and ARC-AGI 2. Available now via API and apps for Pro/Max/Team users, this beast outpaces GPT-5.1 and Gemini 3 Pro while slashing token bills — a paradigm-shifting efficiency bomb in the AI arms race.

Microsoft Research dropped Fara-7B on November 24, 2025 — its first agentic small language model (SLM) built for "computer use," turning screenshots into seamless mouse/keyboard symphonies. With just 7B parameters, this multimodal beast predicts thoughts, clicks, and scrolls with pinpoint accuracy, outperforming UI-TARS and rivaling GPT-4o on benchmarks like WebVoyager (85% success) while averaging 16 steps per task vs. 41 for giants. Open-sourced under MIT on Hugging Face and Azure Foundry, it's primed for on-device Copilot+ PC runs — early tests show devs automating web marathons in sandboxed bliss, no cloud crutches needed.

Meta’s WorldGen: The Generative AI That Spawns Playable 3D Worlds From Text The metaverse's content drought just got irrigated by a generative firehose. Meta’s WorldGen isn’t another toy that spits out pretty 3D blobs you can’t touch — it’s a pipeline powerhouse that births fully playable worlds from bare words, complete with physics, navigable paths, and production-ready pixels that load strai...