Tencent

Youtu-VL-4B-Instruct-GGUF v1.0

Youtu-VL-4B-Instruct-GGUF is a lightweight yet robust Vision-Language Model (VLM) developed by Tencent. Built upon the Youtu-LLM with 4 billion parameters, it introduces the Vision-Language Unified Autoregressive Supervision (VLUAS) paradigm, enhancing visual perception and multimodal understanding. This model excels in both vision-centric and general multimodal tasks without the need for task-specific modules.

CL-bench v1.0

CL-bench is a benchmark designed to evaluate language models' ability to learn from complex, context-dependent tasks. It comprises 500 complex contexts, 1,899 tasks, and 31,607 verification rubrics, all crafted by experienced domain experts.

TencentDB Agent Memory vLatest (May 2026)

TencentDB Agent Memory is a fully local, open-source long-term memory system for AI agents built by Tencent. Using a 4-tier progressive pipeline with symbolic short-term memory (Mermaid canvas) and layered long-term memory (L0→L3 persona hierarchy), it cuts token usage by 61% and improves agent task success by 51% when integrated with OpenClaw — with zero external API dependencies.

WorkBuddy Bench v1.0

WorkBuddy Bench is a multi-domain evaluation suite designed to test coding agents on realistic, complex tasks across software engineering, web development, office workflows, and security. It features 260 contamination-resistant tasks reverse-engineered from real-world commits and business scenarios to ensure agents are evaluated on genuine problem-solving capabilities rather than memorized data.

AngelSpec v0.1.0

AngelSpec is a unified, torch-native training framework designed for speculative decoding, supporting both autoregressive Multi-Token Prediction (MTP) and block-parallel drafting architectures. It enables independent scaling of inference and training through a disaggregated architecture, facilitating high-performance model deployment.

HY-1.8B-2Bit v1.0

HY-1.8B-2Bit is a 2-bit quantized large language model developed by Tencent's AngelSlim project. It utilizes Quantization-Aware Training (QAT) on the Hunyuan-1.8B-Instruct backbone, achieving high performance with significantly reduced model size.

High Performance LLM Inference Operator Library

The High Performance LLM Inference Operator Library is an open-source project developed by Tencent, designed to optimize the inference performance of large language models (LLMs). It provides a set of operators and tools that enhance the efficiency and scalability of LLM deployments, making it easier for developers to integrate and utilize LLMs in various applications.

Cube Sandbox by Tencent v0.1.0

Cube Sandbox is an open-source high-performance sandbox runtime for AI agents built on RustVMM and KVM. It achieves sub-60ms cold starts, under 5MB memory per instance, hardware-level kernel isolation, and supports thousands of concurrent sandboxes per node. 100% E2B SDK compatible with zero code changes to migrate.

HY-MT1.5 v1.5

HY-MT1.5 is a multilingual machine translation model developed by Tencent, available in 1.8B and 7B parameter versions. It supports translation across 33 languages and 5 ethnic and dialect variations, optimized for both on-device and cloud deployment.

Hy3 v3

Hy3 is a powerful 295B-parameter Mixture-of-Experts (MoE) large language model developed by Tencent. It features 21B active parameters per token, a 256K context window, and is designed to deliver high-performance reasoning, agentic workflows, and long-context processing at lower compute costs.