Tencent Hy

Tencent's 3D creation pipeline opens to everyone with no application required. PartGen 1.5 raises component partitioning to 1536³ and adds brush-based manual editing, while HY 3D 3.1 accepts up to eight input views and improves texture fidelity.

Tencent ships Hy3: a 295B-parameter MoE (21B active, 256K context) under Apache 2.0, with a two-week free API and day-one support in Cline, OpenCode, Kilo Code and OpenClaw. 1-bit and 4-bit GGUF builds run the flagship on a single GPU through llama.cpp with MTP.

An 80B MoE (13B active) that reasons over an input image before editing it, using a native chain-of-thought schema trained with Tencent's MixGRPO. Targets add/remove/modify edits that leave untouched regions intact, plus multi-image fusion. The web demo is PC-only.

A speech recognition preview built on Hy3's language model rather than a standalone acoustic stack, which Tencent credits for context awareness and dialect robustness on messy real-world audio. The headline self-reported figure is roughly 3% multilingual word error rate.

MMAE tests whether audio models can edit rather than generate: change what an instruction names and leave everything else untouched. Across 2,000 samples and 17,741 rubric items spanning speech, music and mixtures, exact-match rates sit below 5%, with multi-hop and multi-round edits the weakest.

Tencent open-sources WorldMirror, a universal feed-forward 3D reconstruction model that accepts video, multi-view images, text or a single frame and optionally camera poses, intrinsics and depth as geometric priors. One pass on one GPU yields point clouds, depth maps, normals, camera parameters and

Tencent and Renmin University open-source PlanningBench: 30+ real-world planning tasks with automated verification, usable for both evaluating and training LLM planning, with reported results for current frontier models.

Tencent introduces Hyra-1.0, a research agent built to recursively improve its own solutions on performance-driven research and engineering tasks, with demos published across AI4AI, AI4Science and AI4Fun.

AngelSpec covers drafter training and deployment for speculative decoding. On Hy3-A21B its DFly drafter reports a 1.98–2.40x end-to-end speedup over autoregressive decoding at concurrency 4–64, and 10.5–11.8% higher throughput than DFlash. Drafter weights included.

HunyuanOCR arrives as a 1B-parameter model claiming state-of-the-art text recognition, small enough to run cheaply in a document pipeline. Weights, a hosted demo, web and mobile project pages and a technical report are all public.

HunyuanVideo 1.5 is an 8.3B DiT video model generating 5–10 second 480p/720p clips on consumer GPUs with 14GB of VRAM, with super-resolution to 1080p. Weights, code and the technical report are public.

Tencent and Tsinghua use test-time training to hold fast weights as compact spatial memory over long video, reporting sublinear memory growth past 7,000 frames, 40%+ less compute, and SOTA on VSI-Bench. Paper, code, models and data are all public.

Two translation specialists, a 1.8B dense model and a 30B-A3B MoE, are now routable through OpenRouter, covering 33 languages with instruction following aimed at structured, contextual, and style-guided translation rather than raw sentence pairs.

The Hy-MT2 translation model weights move to Apache 2.0, clearing commercial use, fine-tuning and derivative distribution. Two variants are currently sitting near the top of the Hugging Face trending list.

Tencent open-sources a streaming video diffusion model that generates explorable 3D worlds from text or an image at 24 FPS, holding geometric consistency over long horizons by rebuilding context from past frames rather than letting memory decay. Keyboard and mouse control use a dual action represent

Four new results from Tencent's math agent on open problems: the 3D Blaschke–Lebesgue lower bound rises to 97.9% of the conjectured Meissner optimum, the Beurling–Ahlfors coefficient tightens to 1.523958, and commutator approximation cost drops from O(log⁵(1/ε)) to O(log³).

Two Tencent open-source drops landed at once: HunyuanImage 3.0, an 80B-parameter text-to-image model, and Hunyuan3D-Part, which pairs P3-SAM native 3D part segmentation with X-Part for controllable shape generation. Both are free to deploy.

Tencent's newest flagship preview becomes selectable in OpenCode Go: 770B total parameters with 49B active per token and a 1M-token context window, positioned specifically for coding agents rather than chat.

Tencent opens its 3D generation platform and API internationally: text, multi-view image or sketch to 3D, OBJ/GLB output for Unreal, Unity and Blender, 20 free generations a day for new creators and 200 API credits at signup. HY 3D Studio goes global alongside it.

Tencent open-sources Hy3 preview, a 295B/21B-active reasoning and agent model pitched on cost efficiency, with weights on Hugging Face and a time-boxed free window on OpenRouter while the team collects feedback ahead of the official release.

HPC-Ops is a production LLM inference operator library written in CUDA and CuTe, reporting 30% QPM gains on Tencent's own models and 17% on DeepSeek, with kernel speedups up to 2.22x over FlashInfer/FlashAttention, 1.88x over DeepGEMM and 1.49x over TensorRT-LLM on FusedMoE.

The interactive world model gets the pieces that make it usable outside a demo: training code you can point at your own data, a lite 5B checkpoint that fits small-VRAM GPUs, inference and memory optimizations for real-time interaction, and an open hosted app with no waitlist.

Tencent maps self-evolving agents onto an L0–L4 taxonomy and pairs it with a reliability ladder for deciding when a self-modification should be accepted, backed by an open catalog of 549 works. The core rule: an update must not control the evidence used to accept it.

Tencent's foundation model moves to a 406B/32B-active MoE with a 256K window, split into Think and Instruct variants on Tencent Cloud. The agentic jump is the headline: SWE-bench Verified 6.0→53.0 and Tau2-Bench 17.1→72.4.

An open-weight 770B MoE with 49B active parameters and a 1M-token window, shipped alongside mixed-bit GGUFs that compress 1.5TB to roughly 200GiB by letting calibration data pick each layer's bit-width, reaching 1.31-bit in places and 2.06-bit elsewhere.

Free access to Hy3 preview on OpenRouter has ended and the model reverts to paid pricing. Tencent reports it took 15.4% of provider token volume over the two weeks, with the top position in coding and tool-calling usage.

Hy4 preview becomes selectable in Cline, with 770B parameters, 49B active, a 1M-token context and an Apache-2.0 license. Cline reports it leading SWE-bench Pro and calls it Tencent's largest measured generational jump; setup is npm install, /model, select.

Tencent opens the 2B member of an embodied-agent foundation family aimed at spatial-temporal perception, prediction and planning, with a 32B sibling for heavier reasoning. Notable choices: a Mixture-of-Transformers split by modality, latent perception tokens, and on-policy distillation from the larg

A reasoning-oriented vision-language model, placed third on the arena leaderboard at announcement and callable through Tencent Cloud immediately, with the model release and technical report scheduled for late October.

A billion-parameter flow-matching DiT turning text into 3D character animation, released with weights, code and a technical report. Tencent claims the first motion model trained through a complete pretrain→SFT→RL loop, spanning 200+ motion categories.

GradLoc uses distributed binary search to pin a gradient spike to the exact culprit token in O(log N) with negligible overhead. It surfaced layerwise gradient heterogeneity, specifically tokens with safe importance-sampling ratios that still blow up individual layers, a problem addressed by LayerClip, a per-layer clipping method.

Tencent open-sources UniRL, an RL post-training stack running one generate-score-advantage-update loop across flow/diffusion models, LLMs, VLMs and unified autoregressive+diffusion models. Ships FSDP2 sharding, pluggable SGLang/vLLM-Omni rollouts, and two new algorithms: Flow-DPPO and DRPO.

Tencent released HunyuanImage 3.0-Instruct under open weights, publishing full and distilled checkpoints on Hugging Face plus a GitHub repo. The lab claims top-tier standing on Arena's Image Edit leaderboard for open-source image-to-image editing.

Quantization-aware training takes a 1.8B model to a 2-bit representation, roughly a 0.3B bit-equivalent footprint at 600MB, while keeping dual chain-of-thought reasoning. Reported prefill speeds are 3-8x faster on Apple M4 and Dimensity 9500, with 2-3x faster generation. The model ships as GGUF.

A training-ready 3D corpus: 252k+ filtered high-fidelity objects, 240k+ part-level decompositions for fine-grained control, and 125k synthesized assets for class balance, shipped with Hunyuan3D-2.1-Small as a lightweight reproducible baseline.

An agentic pipeline that composes large-scale open worlds from text prompts and outputs editable, game-ready 3D assets with geometry and textures. The output is explicitly not video frames and not Gaussian splats, so an artist can open the result and modify it.

A 1B-parameter OCR model handling text spotting, table and formula extraction to HTML/LaTeX, video subtitle pull and 14-language photo translation in a single instruction and a single inference pass. It scores 860 on OCRBench among sub-3B models and 94.1 on OmniDocBench.

PolyGen 1.5 lands in Hunyuan 3D Studio with end-to-end native quad mesh generation instead of the tri-meshes earlier generators produced, learning quad topology directly for cleaner, continuous edge loops. Both quad and triangular output are supported.

The 30B-A3B translation model gets an official GGUF build, closing the gap between open weights and actually running them locally on llama.cpp-family runtimes. Tencent also names support across Apple MLX-LM, ONNX Runtime, NVIDIA NeMo and LLaMA-Factory.

Text, image or video in; navigable 3D worlds out as mesh, 3DGS and point clouds that import into Unity and Unreal. The stack is itemized: panorama initialization without camera metadata, VLM-plus-navmesh trajectory planning, latent keyframe expansion, and an engine-agnostic renderer.

Tencent open-sources Hy-MT2 translation models at 1.8B, 7B and 30B-A3B across 33 languages. The 1.8B variant, quantized to 1.25-bit with AngelSlim, fits in 440MB and runs locally on mobile chips at 1.5x the speed of Hy-MT1.5.

CL-bench measures whether models can apply knowledge that exists only in their context: 1,899 expert-curated tasks over 500 complex contexts, gated by 31,000+ validation rules. Ten frontier models complete 17.2% on average, which the authors call a bottleneck for real context-dependent work.

Tencent open-weights a translation pair sized for two very different deployments: a 1.8B model claiming a 1GB footprint and 0.18s latency per 50 tokens on consumer hardware, and a 7B model positioned against far larger general systems. Weights and code are public.

HY-WU synthesizes instance-conditioned LoRA updates from hybrid image-instruction representations and injects them into a frozen backbone during the forward pass, producing a per-input adapter with no test-time optimization. The premise is that shared LoRAs compromise across objectives while single-domain ones fail to generalize.

Day-zero vLLM support for Hy4-preview, verified on NVIDIA GPUs. The serving notes expose the architecture: 256 routed experts plus one shared, a 1M context where each query attends to just 2048 tokens, and only 21 of 78 layers computing their own sparse index.

A memory plugin for persistent agents such as OpenClaw, built on a six-layer framework with a System1/System2 split and a three-stage consolidation chain. Tencent reports 70%+ fewer stored memories, 45% higher density per memory and 35% fewer tokens on very long contexts.

Two open translation models on a shared lineage: a 1.8B aligned by on-policy distillation for consumer hardware at a 1GB footprint and 0.18s latency for 50 tokens, and a 7B upgrade of Tencent's WMT25 winner. Custom terminology, long-dialogue context and format preservation supported.

Tencent surfaced HyCreator, an agent harness aimed at long video generation. Its end-to-end auto mode targets 10-minute-scale films with no human in the loop, and users can drop into real-time interactive editing at any point. Early access is open.

A step-distilled 480p image-to-video checkpoint for HunyuanVideo 1.5 cuts sampling to 8 or 12 steps, a roughly 75% reduction in end-to-end time and under 75 seconds on a single RTX 4090, with quality held near the original. Four steps trades quality for more speed.

Tencent releases the training code for WorldMirror, the Hunyuan World 1.1 model that turns video into 3D worlds in about a second and maps any input, whether image, video, or 3D prior, to 3DGS, depth, cameras, normals, or point clouds. A hosted demo is live.

An RL post-training framework aimed squarely at interactive world models, released with training code you can point at your own data, rewards or base model, plus a checkpoint. The reported gain is following compositional action instructions and staying coherent over long interactions.

Hy-MT1.5-1.8B quantized to 1.25 bits: 3.3GB at FP16 down to 440MB, roughly 25% smaller and 10% faster than 1.67-bit approaches with no claimed accuracy loss. Covers 33 languages, 5 dialects and 1,056 directions, with an Android demo, weights, code and a paper.

Using the Hyra research agent on Hy3, Tencent reports an explicit construction showing the optimal exponent relating |A+A| to |A-A| is exactly 2, a gap open since a 1969 upper bound, where prior constructions barely passed 1.1. A paper and a formal proof are published.