Tencent Hy3

Overview

Tencent Hy3 is Tencent's open-weight flagship large language model, released on July 6, 2026. Tencent's announcement describes it as a "hybrid fast-and-slow-thinking model built on a Mixture-of-Experts (MoE) architecture" with "295 billion total parameters and 21 billion active parameters," supporting "a context length of up to 256K," and shipping "under the commercially friendly Apache 2.0 license."

Caixin Global frames it as the final model of the Hunyuan 3 generation and an upgrade over the preview released in April, "with improved post-training compute and higher-quality training data." "Hy3" and "Hunyuan 3" are used interchangeably.

The pitch is efficiency rather than scale. Tencent claims Hy3 reaches "intelligence comparable to flagship models with two to five times its parameter scale," and backs the claim with a human evaluation rather than only benchmarks: "We ran a blind evaluation with 270 experts using tasks from their work, and Hy3 scored 2.67/4, outperforming GLM-5.1 at 2.51/4."

A note on naming

There is no Hunyuan 4. Tencent has never released a flagship under that name. The only Tencent model whose name contains "Hunyuan-4" is Hunyuan-4B, a 4-billion-parameter edge model published in July 2025 — unrelated to any flagship, and roughly 1/74th of Hy3's size.

An earlier version of this page described a "Hunyuan 4" flagship with 3D generation capabilities, a LiveSearch engine, and a benchmark table. That model never existed. It has been replaced with sourced material about Hy3.

Capabilities

• Hybrid fast and slow thinking: A reasoning_effort parameter selects among three settings — no_think (the default, direct response), low, and high (deep chain-of-thought). One model covers the full latency range.
• Agentic software engineering: 78 on SWE-bench Verified and 57.9 on SWE-bench Pro.
• Graduate-level science: 90.4 on GPQA Diamond.
• Efficiency at 21B active parameters: A 295B MoE that activates 21B per token, with Tencent claiming parity against models two to five times its size.
• Multi-Token Prediction: A single MTP layer (3.8B parameters) supports speculative decoding at serve time, exposed in vLLM through speculative-config.method mtp.
• Production deployment at scale: Already running behind WorkBuddy/CodeBuddy, Yuanbao, Marvis, ima, WeChat customer service, and Path of Exile: Advent on WeGame.

Technical Specifications

• Hugging Face repository: tencent/Hy3
• Total parameters: 295B
• Active parameters: 21B per token
• Experts: 192, top-8 activated
• MTP layers: 1, carrying 3.8B parameters
• Architecture: Mixture-of-Experts transformer, hybrid fast-and-slow thinking
• Context window: 256K tokens
• License: Apache 2.0
• Thinking control: reasoning_effort — no_think (default), low, or high

Tencent does not publish a knowledge cutoff or a training-token count for Hy3, so this page does not state one.

Use Cases

• Coding assistants: Hy3 powers Tencent's own CodeBuddy, and its 78 SWE-bench Verified score is competitive with much larger open-weight models.
• Autonomous agents: Caixin reports "significantly strengthened autonomous-agent capabilities, achieving a 90% task-completion rate across several of the company's internal applications," alongside a free agent feature for consumers.
• Customer service at scale: Deployed behind WeChat's customer-service assistants, a workload that rewards a cheap 21B-active forward pass over a dense frontier model.
• Cost-constrained self-hosting: Apache 2.0 with 21B active parameters puts frontier-adjacent capability within reach of a far smaller cluster than a 1T-parameter MoE requires.
• Latency-tiered products: no_think for interactive turns, low for the middle ground, reasoning_effort: high for the hard ones — all without swapping models.
• Long-document work: A 256K context window handles most contracts, filings, and technical manuals in a single pass.

Performance / Benchmarks

Reported in the benchmark appendix of Tencent's Hy3 model card. Benchmark names below follow the appendix exactly — several are scoped to a subset, and the scoping matters.

Two of these were previously mislabelled on this page. SkillsBench is not "Skillsbench V1": Tencent's note reads "All models are evaluated via Claude Code on 79 tasks (self-contained subset, excluding multimodal tasks); results are averaged over 3 runs." WildClawBench is not an "overall" score: Tencent evaluates it "using the OpenClaw Harness on the text-only subset (35 queries in total)." Calling either figure a full-coverage result overstates what was measured.

Tencent's appendix also carries a footnote that applies to every number above: "The reasoning effort was set to the highest tier for all models." These are not no_think scores.

Blind expert evaluation

Tencent also reports a human evaluation rather than only automated benchmarks: "We ran a blind evaluation with 270 experts using tasks from their work, and Hy3 scored 2.67/4, outperforming GLM-5.1 at 2.51/4."

These are vendor-reported figures, including the blind evaluation, whose task set Tencent selected. Treat them as an upper bound and validate on your own evaluation set. Note also that DeepSWE (28) and Apex-Agent (25.6) are hard agentic benchmarks where absolute scores across the whole field remain low.

Limitations

• Hosted pricing is quoted in CNY only: Tencent Cloud's rate card prices Hy3 in 元 per million tokens. There is no published USD rate card, and yuan prices should not be converted at assumed parity.
• Weaker on the hardest agentic benchmarks: 28 on DeepSWE and 25.6 on Apex-Agent indicate long-horizon autonomy is not yet a strength.
• 256K context, not 1M: Several competing open-weight flagships — DeepSeek V4, GLM-5.2 — now offer a million tokens.
• No published knowledge cutoff: Ground time-sensitive queries with retrieval or tool use.
• Weights, not full openness: Apache 2.0 covers the weights. Training data and the full pipeline are not published.
• Vendor-reported benchmarks: Including the blind expert evaluation. All published scores were run at the highest reasoning-effort tier, so they do not describe no_think behaviour. Independent replication is limited.
• Ecosystem gravity: Most of the shipped integrations sit inside Tencent's own products; third-party tooling is thinner than for DeepSeek or Qwen.

Pricing & Access

Self-hosting is free. The weights are Apache 2.0 licensed and published at tencent/Hy3 on Hugging Face. Commercial use, modification, and redistribution are permitted.

Hosted access is sold through Tencent Cloud, which publishes a rate card on its TokenHub 模型价格 ("Model Pricing") page, last updated 2026-07-07 (最近更新时间：2026-07-07). Under 在线推理场景 ("online inference scenarios"), Hy3 is listed at:

Hy3 preview is priced separately and tiered by input length:

All figures are 元/百万 tokens — yuan (CNY) per million tokens. Tencent publishes no USD rate card for Hy3, and no conversion is implied here.

Consumer access is available through Tencent's own applications — Yuanbao, ima, and WeChat-based assistants — with Caixin reporting a free AI-agent feature at launch.

Serving: vLLM supports Hy3, including MTP-based speculative decoding via speculative-config.method mtp.

Ecosystem & Tools

• tencent/Hy3 on Hugging Face — Apache 2.0 weights, benchmark table, deployment guides
• Hunyuan portal — Tencent's model hub
• Tencent Cloud Hunyuan — managed API access
• Tencent Cloud TokenHub 模型价格 — the published CNY rate card for Hy3 and Hy3 preview
• CodeBuddy / WorkBuddy — Tencent's coding and workplace assistants, running on Hy3
• Yuanbao, Marvis, ima — consumer assistants built on the model
• vLLM — supported serving path with MTP speculative decoding

Community & Resources

• Tencent launches Hy3 — the official announcement, dated July 6, 2026
• Tencent Launches Upgraded Hunyuan 3 AI Model With Free Agent Feature — Caixin Global
• Hy3 model card
• Tencent AI blog
• Compare with GLM-5.2, DeepSeek V4, Kimi K2.6, and Ling-2.6-1T