Tencent Hy icon

Tencent Hy Hy3

Hy3

Improve production reliability and cost efficiency over Hy3 Preview while retaining a 295B MoE architecture with 21B active parameters and 256K context Deploy Hy3 across Tencent products and Tencent Cloud TokenHub, moving developers from a preview checkpoint to an official current release Combine fast and slow thinking for code, reasoning, long-context, and agent tasks in one open-weight model

Reviewed by ToolWorthy Editors·updated today·Hy3 released 5 days ago

Pricing:Free + from ¥1/per use
Jump to section
No media available

Featured alternatives

Coasty icon

Coasty

MakersClaw icon

MakersClaw

BrowserAct icon

BrowserAct

Timbal AI icon

Timbal AI

Atomic Mail Agentic icon

Atomic Mail Agentic

ChatGPT icon

ChatGPT

Pros & Cons

Pros

  • Official release replaces the prior preview without changing the familiar 295B/21B/256K technical profile.
  • Fast-and-slow reasoning is combined in one model family.
  • Wider Tencent product use and TokenHub availability improve the route from evaluation to deployment.
  • Open weights and fine-tuning materials provide an alternative to API-only closed models.

Cons

  • Official claims of stability and efficiency still need workload-specific validation.
  • Commercial terms are governed by a custom community license rather than a standard OSI license.
  • Provider availability, price tiers, and regional access can differ across Tencent surfaces.
  • The 256K context window remains below some competitors' advertised 1M-token limits.

Overview

Tencent Hy3 is the official July 6, 2026 release that follows April's Hy3 Preview. It keeps the same broad technical profile—295 billion total parameters in a mixture-of-experts architecture, 21 billion active parameters, and a 256K-token context window—but Tencent positions this release as more stable and cost-efficient for practical deployments. The main upgrade is therefore operational rather than a new model family: teams evaluating the preview now have a formal current checkpoint with wider product integration.

Hy3 targets reasoning, coding, long-document work, instruction following, and agent workflows. It is particularly relevant to developers who want an open-weight alternative to closed frontier models but still need hosted access through Tencent Cloud. The weights and training ecosystem remain important for organizations that need self-hosting or fine-tuning rather than a chat-only service.

What's New

A formal release after the preview

The preview gave Tencent product teams and external users an early Hy3 checkpoint. The July release replaces that preview as the current entry in the Hy line. Tencent's announcement emphasizes improved post-training quality, stability, and cost efficiency; it does not present Hy3 as a different parameter scale or a longer-context replacement. Existing users should therefore validate production behavior, latency, and API cost on their own workloads rather than assuming a new architecture.

Deeper Tencent product integration

Tencent lists Hy3 as adopted in WorkBuddy/CodeBuddy, Yuanbao, Marvis, ima, and other products. This matters for teams already working inside Tencent's ecosystem: the model is no longer only a downloadable checkpoint or experimental endpoint, but is being used in multiple user-facing and developer-facing surfaces. API availability on Tencent Cloud TokenHub is part of the formal release, while third-party integrations are expected to arrive progressively.

Fast and slow thinking in one MoE model

Hy3 combines fast and slow reasoning behavior in a single MoE model. For practical users, that means one model family can handle low-latency requests and tasks that need more deliberation without switching to a separately branded reasoning checkpoint. The official release retains the 256K context window, making it suitable for long documents, codebases, and agent traces that exceed ordinary chat inputs.

Availability & Access

Hy3 is available through Tencent's product ecosystem and Tencent Cloud TokenHub. Tencent also publishes model materials through its Hy/Hunyuan channels; licensing and hosting terms should be reviewed before commercial deployment. Availability, pricing, and supported regions can vary between Tencent Cloud, self-hosted weights, and third-party providers. Developers should verify the exact model identifier and quota in TokenHub before migrating a production integration.

Pricing & Plans

Tencent Hy offers product access surfaces, open-weight evaluation under the Tencent Hy Community License, and paid hosted API access. Tencent Cloud's Chinese TokenHub pricing page lists the official Hy3 model at CNY 1 per million input tokens, CNY 4 per million output tokens, and CNY 0.25 per million cached-input tokens. Hy3 Preview remains separately listed with length-tiered pricing, so do not reuse preview prices for the official Hy3 release.

Best For

  • Teams already using Tencent Cloud or Tencent productivity products that need a current supported Hy model.
  • Developers testing open-weight reasoning and coding models with a hosted API fallback.
  • Organizations with long-document, codebase, or agent-trace workloads that fit within 256K tokens.
  • ML teams that need fine-tuning options and can review Tencent's community license.

FAQ

Is Hy3 different from Hy3 Preview?

Hy3 is the official July 2026 release after the April preview. Tencent describes better stability, cost efficiency, and product integration while retaining the 295B total/21B active MoE design and 256K context window.

Where can I access Hy3?

Tencent lists Hy3 in several products and through Tencent Cloud TokenHub. Self-hosting and hosted access have different setup, region, and licensing requirements.

Is Hy3 open source?

Hy3 is open weight under the Tencent Hy Community License. That is not the same as an OSI-approved open-source license, so commercial teams should review the terms before deployment.

What does Hy3 cost?

Hosted TokenHub pricing is token-based and varies by input-length tier and output. Check Tencent Cloud's current pricing page because exchange rates, free periods, and tiers can change.

Version History

Hy3

Current Version

Released on July 6, 2026

+What's new
3 updates
  • Improve production reliability and cost efficiency over Hy3 Preview while retaining a 295B MoE architecture with 21B active parameters and 256K context
  • Deploy Hy3 across Tencent products and Tencent Cloud TokenHub, moving developers from a preview checkpoint to an official current release
  • Combine fast and slow thinking for code, reasoning, long-context, and agent tasks in one open-weight model

Hy3 Preview

Released on April 23, 2026

View Update
+What's new
3 updates
  • Run Hy3 Preview as Tencent's current open flagship LLM, using a 295B-parameter MoE model with 21B active parameters and an added 3.8B MTP layer for faster generation
  • Analyze long codebases, documents and agent sessions with up to 256K tokens of context, supported by 80 transformer layers and grouped-query attention
  • Switch between quick answers and deeper reasoning per task, with reported results of 74.4 on SWE-bench Verified, 54.4 on Terminal-Bench 2.0 and 67.1 on BrowseComp

Hunyuan 2.0

Released on December 5, 2025

+What's new
2 updates
  • Access HY 2.0 in Yuanbao, ima and Tencent Cloud APIs with 406B total parameters, 32B active MoE, 256K context, and stronger reasoning, coding and instruction following
  • Mark the international naming shift from Tencent Hunyuan to Tencent HY, while keeping Hunyuan as the technical family name used in repositories, APIs and earlier documentation

Hunyuan Compact Series (0.5B / 1.8B / 4B / 7B)

Released on July 30, 2025

+What's new
2 updates
  • Deploy compact Hunyuan models in 0.5B, 1.8B, 4B and 7B sizes, with Pretrain and Instruct checkpoints available on Hugging Face for local and edge use
  • Fine-tune smaller Hunyuan models for consumer GPUs, PCs, mobile and embedded environments, while retaining long-context, hybrid-reasoning and agent-oriented capabilities

Hunyuan-A13B

Released on June 25, 2025

+What's new
2 updates
  • Run Hunyuan-A13B as an open MoE model with 80B total parameters, 13B active parameters and 256K context for efficient reasoning, coding and agent workloads
  • Switch between fast and slow thinking modes in a smaller open model, giving developers a lower-cost bridge between T1-style reasoning and later compact dense releases

Hunyuan-T1 (Official)

Released on March 21, 2025

+What's new
2 updates
  • Use Hunyuan-T1 as the official deep-reasoning model built on TurboS, improving complex math, logic, science and coding tasks over the earlier preview version
  • Access T1 through Tencent Cloud and Tencent Yuanbao for long-context reasoning workflows, rather than relying only on the earlier Yuanbao preview experience

Hunyuan TurboS

Released on February 27, 2025

+What's new
3 updates
  • Use Hunyuan TurboS for fast-answer workloads that need lower latency, with public claims of doubled output speed and 44% lower first-token latency versus the prior generation
  • Run long-context and reasoning workloads on the Hybrid-Transformer-Mamba MoE base that later powered Hunyuan-T1, while reducing sequence-processing overhead
  • Build on the TurboS base for adaptive fast and slow reasoning, choosing quicker replies for simple prompts and deeper reasoning for complex tasks

Hunyuan T1-Preview

Released on February 17, 2025

+What's new
2 updates
  • Try Hunyuan T1-Preview in Tencent Yuanbao as Hunyuan's first consumer-facing deep-reasoning model before the official T1 release
  • Test early reasoning behavior on math, coding and complex problem-solving tasks, giving Tencent feedback before the TurboS-based official version arrived in March 2025

Hunyuan-Large (389B MoE)

Released on November 5, 2024

+What's new
3 updates
  • Open-source the largest Transformer-based MoE LLM at the time — 389B total parameters with 52B active parameters and 256K context — under the Tencent Hunyuan Community License Agreement
  • Download Pretrain, Instruct and Instruct-FP8 checkpoints with the technical report, covering math, code, multilingual and long-context benchmark evaluations
  • Outperform Llama 3.1 70B and 405B on multiple academic benchmarks at launch, establishing Hunyuan-Large as a reference open MoE baseline for the research community

Hunyuan Turbo

Released on September 5, 2024

+What's new
3 updates
  • Launch the next-generation Hunyuan flagship at the Tencent Global Digital Ecosystem Summit as the first Hunyuan model to fully adopt a Mixture-of-Experts architecture for production serving
  • Double training efficiency, double inference throughput and cut inference cost by roughly 50% versus Hunyuan 1.0, with input and output API pricing reduced to half of the previous generation
  • Apply the Turbo service to language understanding, text generation, math and code tasks while relying on Tencent's published benchmark comparisons for quality claims

Hunyuan 1.0

Released on September 7, 2023

+What's new
3 updates
  • Access Tencent Hunyuan as Tencent's first proprietary foundation model on Tencent Cloud, launched with more than 100B parameters and over 2T pretraining tokens
  • Connect Hunyuan-backed capabilities across 50+ Tencent products at launch, including Tencent Cloud, Tencent Meeting, Tencent Docs, Weixin Search and QQ Browser
  • Open the model to enterprises in China for testing and app building through Tencent Cloud, marking Tencent's formal entry into the Chinese frontier LLM race

From the blog

View all →

Track Tencent Hy in ToolWorthy Weekly

Important tool updates, better alternatives, and selected AI signals in one weekly brief.

Weekly only. Unsubscribe anytime.