Hy3
Hy3 is Tencent Hy's hybrid fast-and-slow-thinking Mixture-of-Experts model with 295 billion total and 21 billion active parameters and up to 256K context, released with open weights under Apache 2.0. Tencent says it is used across WorkBuddy/CodeBuddy, Yuanbao, Marvis and ima.
Released
July 6, 2026
Type
llm
License
open-weight
Apache 2.0
Context
—
Tencent states a context window of up to “256K” tokens.
Pricing
Input
$0.132 / 1M tokens
Output
$0.528 / 1M tokens
List price on the developer's own pages, checked . Tencent Cloud international TokenHub, Singapore region; cached input $0.033 per 1M tokens.
Capabilities
Architecture
Parameters: 295B total, 21B active (MoE), plus 3.8B MTP layer
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- Tencent press release: Tencent Hunyuan Officially Releases Hy3 (2026.07.06) (opens in a new tab)
- Tencent press release: Tencent Hy3 Now Available Globally (August 5, 2026) (opens in a new tab)
- Hugging Face: tencent/Hy3 model card (opens in a new tab)
- Tencent Cloud international: TokenHub model pricing (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Hy3 and Tencent
MarkTechPost
Perplexity AI Releases pplx-embed-v2-late: A 0.6B Edge Model and a 9B Model Scoring 92.4% on MADQA
Perplexity AI released pplx-embed-v2-late, a new suite of ColBERT-style multimodal embedding models, on an unspecified date, offering two distinct sizes: a 0.6 billion parameter model optimized for speed and cost-efficiency, and a 9 billion parameter model designed for maximum quality. Both models are capable of retrieving text, images, and rendered PDF pages, and importantly, they share a unified embedding space, allowing for cross-modal understanding. These models are available for self-hosting under the MIT license, which permits commercial use, and are accessible on Hugging Face. Perplexity has indicated plans for a hosted API endpoint, though it is not yet live. The 0.6B model is engineered to function as a lightweight query encoder, suitable for deployment on edge devices or laptops, utilizing approximately 340 million active parameters for image processing. This smaller model aims to maintain performance close to larger, 8 billion parameter rivals. A key feature highlighted is the ability to search a 9B index using 0.6B queries, which recovers about half of the quality gap observed in text retrieval at the reduced query cost. The models generate 128-dimensional token vectors, significantly narrower than competitors' vectors which range from 2,048 to 4,096 dimensions, representing a 16x to 32x reduction in dimensionality. However, the models present certain limitations. The architecture stores one vector per token, leading to index sizes that scale directly with document length. In terms of performance benchmarks, pplx-embed-v2-late is not the top performer on the ViDoRe v3 image retrieval task; Tencent's EVIE model reportedly scores higher. Furthermore, a single input cannot simultaneously contain both text and images. Perplexity has stated that all performance scores are self-reported, and the accompanying technical report has not yet been published. Technical specifications reveal that the pplx-embed-v2-late-0.6b model comprises 594 million total parameters, with approximately 240 million active parameters for text and 340 million for images. The base model is derived from Qwen3.5-0.8B, pruned to 12 text layers. The 9B variant has 9 billion total parameters (though Hugging Face lists it as 8B) and utilizes 7.4 billion parameters. Both output 128 dimensions per token. Estimated memory requirements for weights in bf16 precision are around 1.2 GB for the 0.6B model and 16 to 18 GB for the 9B model. The models require sentence-transformers version 6.0.0 or later and transformers version 5.4.0 or later. The published checkpoints are distributed in F32 format, which doubles their download size.
The Hacker News
China-Linked UNC3569 Exploited Sogou Input Method Flaw to Deploy GRAYRABBIT Backdoor
A China-linked cyber-espionage group, identified as UNC3569, has successfully exploited a critical vulnerability within the Sogou Input Method, a widely adopted software for typing Chinese characters on Windows operating systems. This exploitation allowed the group to deploy a sophisticated backdoor, known as GRAYRABBIT, onto the compromised systems, according to research published by Gen Digital on Thursday. The initial vector for this attack involved a meticulously crafted link, which, upon interaction by the victim, initiated a chain of malicious actions. The ultimate outcome of the compromise granted the attackers the same level of access and control as the logged-in user, enabling them to perform any action within the user's privileges. This level of access is highly concerning as it can facilitate extensive data exfiltration, lateral movement within a network, and the deployment of further malicious payloads. Sogou Input Method is a prevalent tool used by millions of users, particularly in China, making this vulnerability a significant threat to a large user base. The company Tencent, which owns Sogou, has not yet issued a public statement regarding the exploit or the specific vulnerability. The GRAYRABBIT backdoor is described as a versatile tool that can be used for various malicious purposes, including surveillance, data theft, and maintaining persistent access to compromised environments. The group's modus operandi suggests a focus on espionage, aiming to gather intelligence from targeted individuals or organizations. The exploitation of input method editors (IMEs) like Sogou presents a unique attack surface, as these tools are deeply integrated into the operating system and often run with elevated privileges. Attackers leveraging such vulnerabilities can bypass traditional security measures more effectively. Gen Digital's research highlights the ongoing sophistication of nation-state-backed hacking groups and their ability to identify and exploit niche software vulnerabilities. The detailed analysis of the attack chain underscores the importance of timely patching and robust endpoint security solutions to mitigate such threats. The implications of this attack extend beyond individual users, potentially impacting businesses and government entities that rely on Sogou Input Method for their operations. The ability of UNC3569 to deploy a backdoor with such extensive privileges raises concerns about the potential for long-term surveillance and data compromise. Further investigation into the specific CVE or vulnerability identifier exploited is crucial for understanding the technical details of the attack and for developing effective countermeasures. The incident serves as a stark reminder of the persistent threat posed by advanced persistent threats (APTs) and the need for continuous vigilance in the cybersecurity landscape.
The Hacker News
WeChat Zero-Click Worm Took Over Accounts on iPhone and Android via Incoming Calls
Security researchers at the firm Calif have developed a sophisticated zero-click worm capable of compromising WeChat accounts on both iOS and Android devices. This worm exploits an incoming call vulnerability, meaning that a user does not need to answer the call or interact with their phone in any way for their account to be taken over. The only prerequisite for the attack is that the malicious caller must already be a contact within the target's WeChat address book. Calif researchers successfully demonstrated the worm's ability to spread autonomously between three test phones, highlighting its potential for widespread exploitation. The security firm reported this critical flaw to Tencent, the parent company of WeChat, in July. Tencent has since addressed the vulnerability, indicating that patches have been implemented to mitigate the risk. However, the exact timeline for the fix and the specific versions of WeChat affected were not detailed in the initial report. The discovery underscores the persistent threat of zero-click exploits, which are particularly dangerous because they require no action from the victim, making them extremely difficult to detect and prevent. Zero-click vulnerabilities are highly sought after by malicious actors and intelligence agencies due to their stealth and effectiveness. They bypass traditional security measures that rely on user awareness or interaction, such as clicking on malicious links or opening infected attachments. The ability of this worm to spread via an incoming call, even to contacts, suggests a deep-seated vulnerability within the WeChat application's handling of network communications or call processing. The fact that it affects both major mobile operating systems, iOS and Android, further amplifies the severity of the discovery. Calif's research team has a history of uncovering significant security vulnerabilities in popular applications. Their work often involves intricate reverse engineering and deep analysis of software protocols. The development of this worm demonstrates a high level of technical expertise and a thorough understanding of how communication applications handle network traffic and incoming signals. The successful demonstration on multiple devices and operating systems provides concrete evidence of the exploit's viability and the potential impact on WeChat's massive user base, which numbers over a billion active users globally. The prompt reporting to Tencent and subsequent patching by the company is a positive outcome, preventing a potentially catastrophic widespread breach.
The Hacker News
18-Year-Old Linux SCTP Flaw Could Let Local Users Gain Root and Escape Containers
A critical use-after-free vulnerability within the Linux kernel's Stream Control Transmission Protocol (SCTP) networking code, present since 2008, has been identified and exploited to grant local users full root privileges on a host system. Tencent researchers demonstrated that this flaw can be leveraged to break out of containerized environments and gain access to the underlying machine. The vulnerability, tracked as CVE-2024-XXXX (specific CVE ID not provided in the source), affects older kernel versions that have not yet been patched. The Linux kernel developers have released fixes for this issue in several stable kernel branches. Specifically, kernel versions 7.1.6, 6.18.42, 6.12.101, and 6.6.148 all contain the necessary patches to address the SCTP flaw. These updates were made available on August 3. Users running older, unpatched kernel versions that have SCTP enabled are strongly advised to update their systems immediately to mitigate the risk of exploitation. SCTP is a transport layer protocol that provides message-oriented, reliable, and ordered delivery of data, and it is often used in telecommunications and network infrastructure. The use-after-free bug occurs when memory is deallocated but a pointer to that memory is still in use, leading to unpredictable behavior and potential security exploits. By chaining this memory corruption vulnerability with other techniques, attackers can achieve arbitrary code execution and elevate their privileges to the highest level on the system. The ability to escape containers is particularly concerning, as it undermines the isolation that containers are designed to provide, allowing malicious actors to potentially compromise the host operating system and any other containers running on it. The long-standing nature of the bug, dating back to 2008, highlights the persistent challenge of identifying and rectifying deep-seated vulnerabilities in complex software like the Linux kernel. This discovery underscores the importance of continuous security auditing and timely patching of operating system components, especially for systems exposed to network traffic or running containerized workloads. The fix involves careful memory management within the SCTP implementation, ensuring that memory is properly freed and pointers are invalidated to prevent the use-after-free condition. The disclosure by Tencent researchers serves as a stark reminder that even mature and widely-used software can harbor significant security risks that may be exploited by sophisticated attackers. System administrators and security professionals should prioritize verifying their kernel versions and applying the recommended updates to safeguard their infrastructure against this critical vulnerability.
MarkTechPost
Tencent Open-Sources AngelSpec: A Unified Training Framework for MTP and Block-Parallel Speculative Decoding on Hy3 Models
Tencent has open-sourced AngelSpec, a torch-native training framework designed to enhance speculative decoding for large language models. This framework specifically addresses the challenges of workload heterogeneity by offering specialized training for both autoregressive multi-token prediction (MTP) and block-parallel speculative decoding, particularly for Hy3 models. Speculative decoding is a technique that aims to accelerate text generation by using a lightweight "drafter" model to propose multiple future tokens, which are then verified by a larger "target" model in a single forward pass. The efficiency of this process hinges on two factors: the acceptance rate of the proposed tokens and the computational cost of the draft-and-verify step. These factors often exhibit an inverse relationship across different types of tasks. In open-ended conversational scenarios, where many continuations are semantically plausible, the acceptance rate of draft tokens tends to decrease rapidly as the proposed sequence length increases. This makes generating and verifying long blocks computationally wasteful. For such tasks, AngelSpec employs an autoregressive MTP drafter that proposes shorter candidate sequences, aligning better with the characteristics of conversational data. Conversely, workloads involving code generation or mathematical reasoning exhibit stronger constraints on future tokens due to factors like programming syntax, formal expressions, and step-by-step derivations. These tasks often involve longer predictable spans, which are more effectively handled by block-parallel drafting methods. To cater to this diversity, AngelSpec provides two distinct drafters: an MTP model trained on conversational data and a block-diffusion model optimized with code and mathematics-focused samples. This dual approach allows AngelSpec to tailor the speculative decoding process to the specific demands of different workloads, moving beyond a one-size-fits-all universal drafter. The framework's design emphasizes treating workload heterogeneity as a fundamental constraint, influencing the structure, training data, and verification depth of the models. This contrasts with prior speculative decoding research, which often focused on a single drafter optimized for an averaged benchmark mixture. AngelSpec's commitment to lossless speculative decoding means that the acceleration achieved does not come at the cost of output accuracy. The original Hy3 model, which AngelSpec supports, is trained with a single MTP layer and without recurrent self-conditioned unrolling, with the block being reusable during inference. This open-source release by Tencent aims to provide researchers and developers with a more adaptable and efficient tool for improving the performance of large language models across a wider range of applications.