Verdict: Tencent's Hy3 is a 295-billion-parameter mixture-of-experts model that activates only 21B parameters per token, making it one of the most cost-efficient frontier-class AI agent models available in 2026. Released under Apache 2.0 on July 6, 2026, and available free on Tencent's WorkBuddy workspace until August 31, 2026, it is the fastest way to try a production-grade Chinese AI agent at zero cost. It matches or exceeds GPT-5.5 on several science and search benchmarks, runs workflows of up to 495 sequential steps in production, and topped OpenRouter's global LLM usage leaderboard within its first week.
Last verified: 2026-08-07 · Best for: trying a frontier-class AI agent free before the August 31 cutoff · Best free access: WorkBuddy (global, no card required) · Best for self-hosting: Apache 2.0 weights on HuggingFace · Volatile facts: pricing and free-access windows change often — re-check before committing.
What is Tencent Hy3 and why does it matter in 2026?
Tencent Hy3 (also called Hunyuan 3.0) is Tencent's flagship open-weight AI model, officially released on July 6, 2026, by the Tencent Hy Team (formerly Tencent Hunyuan). It uses a Mixture-of-Experts (MoE) architecture with 295 billion total parameters but activates only 21 billion per forward pass — roughly 7% of its full capacity per token. This means you get the intelligence of a very large model at a fraction of the compute cost, because the model dynamically routes each token to the 8 most relevant "experts" out of 192 available specialist sub-networks (Tencent official announcement, July 6 2026; GitHub: Tencent-Hunyuan/Hy3).
The model was rebuilt from scratch in under six months. In late January 2026, Tencent tore down its entire pre-training and reinforcement-learning infrastructure and started over. The preview launched April 23, 2026. The full release landed July 6, 2026. That compressed timeline — infrastructure rebuild to flagship release in under six months — is itself notable in a field where major model development cycles typically span a year or more (Tencent official announcement; i-scoop.eu analysis, July 6 2026).
Hy3 was co-designed alongside the teams running Tencent's own products — WorkBuddy, CodeBuddy, Yuanbao, ima, and WeChat — and was trained against millions of real user workflows before going public. Chief AI Scientist Yao Shunyu, a former OpenAI researcher who joined Tencent, said the team focused on three things: broad capability across real tasks, evaluation against live product use cases instead of synthetic benchmarks, and cost efficiency built in from day one (Tencent official announcement; NationPress, August 5 2026).
How does Hy3 actually perform on benchmarks?
Hy3 posts competitive results across reasoning, search, and agentic tasks, though it trails frontier models on the hardest coding and pure-math benchmarks. Tencent reported these vendor-verified scores for the official July 2026 release:
| Benchmark | Hy3 Score | What It Measures |
|---|---|---|
| BrowseComp | 84.2 | Web search and research accuracy — matches GPT-5.5 (~84.4) |
| GPQA Diamond | 90.4 | Graduate-level science reasoning |
| USAMO 2026 | 72.0 | Math Olympiad-level reasoning |
| Frontier Science Olympiad | 74.8 | Scientific reasoning — surpasses GPT-5.5 (73.8) |
| HLE (with tools) | 53.2 | Hard language evaluation — ahead of DeepSeek V4 Pro |
| SWE-bench Verified | 78.0 | Real GitHub issue fixes |
| Expert blind eval | 2.67 / 4 | 270 experts, 312 real workflow comparisons |
(GitHub: Tencent-Hunyuan/Hy3; VentureBeat, July 2026; overcentral.com; i-scoop.eu)
Where Hy3 wins: Agentic search (BrowseComp 84.2 is level with Claude Opus 4.8 at 84.3 and GPT-5.5 at 84.4), scientific reasoning (it tops Frontier Science Olympiad at 74.8, ahead of both GLM-5.2 and GPT-5.5), and token efficiency. In the blind expert evaluation, Hy3's biggest advantage over GLM-5.1 was in frontend development and data tasks. Tencent's research blog also notes that Hy3 completed WorkBuddy document tasks using 47.4% fewer tokens than the earlier GLM-5 model, which means lower cost per task (hy.tencent.com/research/hy3).
Where Hy3 trails: On the hardest coding benchmarks, GLM-5.2 (a 744B MoE with ~40B active) leads: SWE-bench Verified 84.2 vs Hy3's 78.0, Terminal-Bench 2.1 81 vs 71.7. On pure math at the elite level, GPT-5.5 leads MathArena Apex 85.4 vs Hy3's 38.7. Claude Opus 4.8 still leads SWE-bench Pro at 69.2 vs Hy3's 57.9 (VentureBeat; i-scoop.eu). The honest picture: Hy3 trades the absolute coding/math crown for cost efficiency and agentic reliability at roughly half the active parameters of its closest competitors.
What is WorkBuddy and how do you get Hy3 for free?
WorkBuddy is Tencent's AI agent workspace — a desktop application (and mobile/IM integration) that takes a natural-language instruction, autonomously plans the task, calls tools, generates files, and delivers a finished result you can verify. It is not a chatbot; it is an agentic workspace built for multi-step office work: research, document creation, data analysis, spreadsheets, and browser-based tasks (Tencent Cloud WorkBuddy product page; workbuddy.ai).
As of August 5, 2026, Tencent expanded global access to Hy3 through WorkBuddy, making it free for users worldwide until August 31, 2026. No credit card is required. You download the desktop client from workbuddy.ai, create an account, and can immediately assign tasks to Hy3-powered agents (Tencent global rollout announcement; NationPress, August 5 2026).
WorkBuddy also supports switching between other Chinese models — DeepSeek, GLM, Kimi, and MiniMax — so you can compare Hy3 against alternatives on the same task without leaving the workspace. This matters for practical evaluation: the best model for a given workflow often depends on the specific task type, not raw benchmark scores. If you want to understand the broader landscape of AI agents that do real work, our guide to 10 AI agents that do real-world work in 2026 covers the full field.
What did Hy3 achieve in production at Tencent?
Real-world deployment metrics from Tencent's own WorkBuddy and CodeBuddy production environments tell the most compelling story. After integrating Hy3:
| Metric | Improvement | Source |
|---|---|---|
| Task resolution rate | 72% → 90% | Tencent research blog |
| Latency | Dropped 54% | Tencent, July 2026 |
| End-to-end task time | Dropped 47% | Tencent, July 2026 |
| Max stable agent workflow | 495 sequential steps | Tencent, July 2026 |
| Daily token consumption | Up 20x since preview | Tencent, July 2026 |
| WorkBuddy users actively choosing Hy3 | Grew 6x | Tencent, July 2026 |
| Hallucination rate | 12.5% → 5.4% | GitHub model card |
| Commonsense error rate | 25.4% → 12.7% | GitHub model card |
| Multi-turn issue rate | 17.4% → 7.9% | GitHub model card |
(Tencent official announcement; GitHub: Tencent-Hunyuan/Hy3; hy.tencent.com/research/hy3; i-scoop.eu)
The 495-step number is the most striking. A workflow of 495 consecutive sequential steps running stably in production is not a demo-environment result — it means the model can execute extended, multi-tool agent chains without breaking, hallucinating, or losing the thread. For anyone building multi-step automated workflows, this reliability is the difference between a toy and a tool. If you are building multi-agent systems, our guide on how to build a multi-agent AI team your whole company actually uses pairs well with Hy3 as a backbone model.
How much does Hy3 cost after the free period?
After the free WorkBuddy window closes on August 31, 2026, pricing falls into two tiers:
Self-host (Apache 2.0 — free): Download the weights from HuggingFace (tencent/Hy3 or tencent/Hy3-FP8 for the quantized version) and deploy with vLLM or SGLang on your own GPUs. Recommended hardware: 8x H20-3e or equivalent high-memory GPUs for the full BF16 model. No per-token cost — you pay only for your own infrastructure (GitHub: Tencent-Hunyuan/Hy3).
Tencent Cloud API (per million tokens):
| Price (RMB) | Approx. USD (~7.2 RMB/$) | |
|---|---|---|
| Input | ¥1.00 / M tokens | ~$0.14 |
| Output | ¥4.00 / M tokens | ~$0.56 |
| Input (cache hit) | ¥0.25 / M tokens | ~$0.03 |
(hy.tencent.com/research/hy3; KuCoin, July 13 2026)
WorkBuddy subscription (after free tier):
| Plan | Price | Credits |
|---|---|---|
| Free | $0 | 5,000 credits (one-time) |
| Personal Pro | ~$9.95/month | 2,000 credits/month |
| Team | ~$40/seat/month | Shared pool + admin console |
(MakerStack review, July 2026; Eigent.ai review, June 2026)
At roughly $0.14/$0.56 per million input/output tokens, Hy3 is among the cheapest frontier-class API options in 2026. The cached-input rate of ~$0.03/M tokens makes repeated queries on the same context dramatically cheaper. For comparison, this is well below the pricing of most closed-fronterier models, which is precisely why it topped OpenRouter's usage leaderboard — the platform is a price-performance vote, and Hy3 wins on cost-to-capability ratio (aimodeling.com, July 16 2026).
Why did Hy3 top the OpenRouter leaderboard?
Within its first week of launch, Hy3 recorded 68 times as many API calls as the previous-generation Hunyuan model and ranked #1 on OpenRouter's global LLM usage leaderboard. Tencent confirmed this in its global rollout announcement on August 5, 2026 (Tencent global rollout; aimodeling.com).
The free access window on OpenRouter (through July 21, 2026) drove initial adoption — Hy3's free version processed 6.13 trillion tokens in a single week on OpenRouter, surpassing Xiaomi's MiMo-V2.5 (5 trillion) and DeepSeek V4 Flash (KuCoin, July 13 2026). But the sustained usage after the free window — and WorkBuddy users growing 6x — signals that developers are returning because the model delivers, not just because it's free. This is the "call volume is the new benchmark" thesis: when developers can choose any model, the one they call most is the one with the best price-to-capability ratio for real work.
How to put Hy3 to work for your business (3 practical workflows)
The fastest way to evaluate Hy3 is not to read about it — it is to run a real task and inspect the output. Here are three workflows you can run today on WorkBuddy with Hy3, each designed to save measurable time:
1. Research-to-content pipeline (one prompt → a week of posts)
Give Hy3 this instruction in WorkBuddy:
"Search for the top 10 questions [your audience] is asking about [your topic] right now. Identify the five biggest frustrations. Write five [LinkedIn posts / blog intros / newsletter sections] that answer each one and position [your business] as the place where they go to solve these problems. Make each post direct, useful, and end with a clear reason to learn more."
Hy3 will browse the web, read what people are actually asking, and generate content grounded in real conversations rather than generic guesses. Because it has a 256K context window, you can paste in your existing content, brand voice document, and previous posts to match your tone across all five outputs. What comes back is five pieces of content, each rooted in a real conversation — enough for a week.
2. Multi-day email nurture sequence
"Write a 7-day email sequence for someone who just opted in to learn about [your topic]. Day 1: welcome and explain what [topic] is in plain language. Days 2–6: one new use case per email with a real-world example of how someone is using it to save time. Day 7: explain what's inside [your offer] and why joining now matters. Keep every email under 300 words. Simple language, no jargon."
Paste your existing welcome emails and brand voice guide into the context window first. Hy3 will match your tone across all seven emails without you rewriting a word. This is the kind of multi-step deliverable that a chat-only tool cannot do in one shot — WorkBuddy's agent architecture means it plans the sequence, drafts each email, checks the word count, and assembles the full set.
3. Research brief + landing page draft
"Research the current landscape of [your industry/product category]. Produce: (1) a 2-page competitive analysis covering the top 5 alternatives, their pricing, and their key strengths/weaknesses; (2) a landing page brief that positions [your product] against those alternatives, with a headline, 3 key benefits, a comparison table, and a call-to-action."
WorkBuddy will output both a formatted document and the landing page brief — files you can immediately hand to a designer or developer. The 256K context window means you can include your product details, pricing, and existing marketing copy for Hy3 to reference.
For teams evaluating open-source AI agent tools more broadly, our guide to open-source AI agent tools on GitHub that are quietly changing how developers work covers Hy3 alongside other notable projects.
What does Hy3's architecture mean for cost and performance?
The Mixture-of-Experts design is the key to Hy3's value proposition. Here is what the numbers mean in practice:
- 295B total, 21B active = roughly 14:1 sparsity. Each token is processed by only ~7% of the model. You get the knowledge capacity of a 295B model with the inference cost closer to a dense 20B model (vast.ai model library).
- 192 experts, top-8 routing = each forward pass activates 8 of 192 specialist sub-networks. The routing mechanism decides which experts are relevant for each token, so simple queries use cheaper experts while complex problems activate deeper reasoning chains (GitHub model card).
- 256K context window = enough for most agent and document tasks, though it trails the 1M-token windows of DeepSeek V4 and GLM-5.2 for very long single-request work (aitraining2u.com).
- 3.8B MTP (multi-token prediction) layer = a small auxiliary layer on top of the 80-layer main network that predicts multiple tokens at once, speeding up generation.
- Hybrid fast/slow thinking = the model can switch between fast pattern-matching mode (
reasoning_effort: "no_think") for simple responses and deep chain-of-thought (reasoning_effort: "high") for complex problems, controlled at inference time (GitHub quickstart).
This architecture also matters for how LLM quantization and model compression work in 2026 — the FP8 quantized version (tencent/Hy3-FP8) makes self-hosting more accessible by reducing VRAM requirements while preserving most of the model's capability.
What are the limitations and caveats?
- Coding is not its strongest suit. GLM-5.2 leads on SWE-bench Verified, Terminal-Bench, and DeepSWE. If your primary use case is autonomous coding, Hy3 is a solid option but not the top open model. Developers may find dedicated coding agents stronger — see our guide comparing Qwen 3.8 Max vs Claude Fable 5 vs GPT-5.6 Sol for task-specific routing.
- 256K context, not 1M. For extremely long single-document analysis, DeepSeek V4 and GLM-5.2 offer 1M-token windows. For most agent and document workflows, 256K is ample.
- Western data sovereignty. WorkBuddy runs on Tencent Cloud infrastructure based in China. Teams in heavily regulated industries with strict data residency requirements should evaluate Tencent's compliance posture carefully before processing sensitive data (Eigent.ai review).
- Vendor-reported benchmarks. The production metrics (72% → 90% task resolution, 54% latency drop, 495 steps) come from Tencent's internal deployment. They are directionally credible but should be validated against your own workloads before committing to the model for production.
- Credit metering on WorkBuddy. The free tier includes 5,000 credits, which deplete quickly on heavy multimodal or long-running tasks. Monitor usage if you are running Hy3 through WorkBuddy rather than self-hosting (AICost review).
- Free window ends August 31, 2026. After that, WorkBuddy transitions to paid tiers. The self-hosted Apache 2.0 weights remain free indefinitely.
What this means for you
If you are a small business owner, solo operator, or team lead evaluating AI agents in 2026, the Hy3 + WorkBuddy combination is the lowest-friction way to test a frontier-class Chinese AI agent today. Three actions to take before the free window closes:
- Download WorkBuddy from workbuddy.ai and run one real task — a research brief, a content pipeline, or a data analysis job. Inspect the output quality, not the benchmark numbers.
- Compare Hy3 against the model you currently use on the same task. WorkBuddy lets you switch between Hunyuan, DeepSeek, GLM, Kimi, and MiniMax — run the identical prompt on two models and compare.
- If the output meets your bar, download the weights from HuggingFace and evaluate self-hosting for cost control and data sovereignty. The Apache 2.0 license means you can use it commercially with no restrictions.
The broader signal: Chinese open-weight models are no longer just competitive on price — they are now competing on capability and real-world reliability, and they are free to try. That changes the calculus for any team currently paying premium API rates for closed-frontier models.
FAQ
Q: Is Tencent Hy3 really free?
A: Yes, Hy3 is free on WorkBuddy for global users until August 31, 2026 (no credit card required). The model weights are also free to download and self-host under the Apache 2.0 license, which has no expiration. After the free WorkBuddy window, API access costs approximately ¥1 ($0.14) per million input tokens and ¥4 ($0.56) per million output tokens on Tencent Cloud. (Tencent global rollout; hy.tencent.com/research/hy3)
Q: How big is Tencent Hy3?
A: Hy3 has 295 billion total parameters but activates only 21 billion per token using a Mixture-of-Experts architecture with 192 experts (top-8 routing). It has a 256K-token context window and an additional 3.8B-parameter multi-token-prediction layer. This means it has the knowledge capacity of a very large model but runs at the cost of a much smaller dense one. (GitHub model card; Tencent announcement)
Q: Can I use Hy3 commercially?
A: Yes. Hy3 is released under the Apache License 2.0, which permits commercial use, modification, and distribution with no restrictions. Businesses can download the weights, fine-tune them, deploy them internally, and use them in commercial products without licensing fees. (GitHub LICENSE)
Q: Is Hy3 better than GPT-5.5?
A: On some benchmarks, yes: Hy3 matches GPT-5.5 on BrowseComp (84.2) and surpasses it on Frontier Science Olympiad (74.8 vs 73.8) and GPQA Diamond (90.4). However, GPT-5.5 leads on the hardest math benchmarks (MathArena Apex 85.4 vs Hy3's 38.7) and some coding tasks. Hy3's advantage is cost efficiency — it delivers comparable agentic and reasoning performance at roughly half the active parameters. (VentureBeat; i-scoop.eu)
Q: What is the difference between Hy3 and WorkBuddy?
A: Hy3 is the AI model (the "brain"). WorkBuddy is the AI agent workspace (the "hands" — the application that uses Hy3 to plan tasks, call tools, browse the web, and generate files). You can use Hy3 without WorkBuddy by deploying it via vLLM or SGLang, or through the API. WorkBuddy simply provides the agentic runtime that makes Hy3 easy to use for non-technical users. (workbuddy.ai; Tencent Cloud WorkBuddy)
Q: How do I self-host Hy3?
A: Download the weights from HuggingFace (tencent/Hy3 for BF16 or tencent/Hy3-FP8 for the quantized version). Deploy with vLLM or SGLang on a server with 8x H20-3e GPUs (or equivalent high-memory cards). The model exposes an OpenAI-compatible API, so you can drop it into any existing toolchain that supports the OpenAI API format. Full deployment instructions are in the GitHub repository quickstart. (GitHub: Tencent-Hunyuan/Hy3)
Q: What hardware do I need to run Hy3 locally?
A: You do not need a powerful computer to use Hy3 via WorkBuddy — the model runs in Tencent's cloud and your machine is just the client. To self-host, you need a server with 8x H20-3e GPUs or equivalent for the full BF16 model. The FP8 quantized version (tencent/Hy3-FP8) reduces VRAM requirements and is more accessible for teams with smaller GPU budgets. (GitHub quickstart; aitraining2u.com)
Every claim here is traced to a primary source, dated, and listed under Sources. Research and drafting are AI-assisted; editing, verification and publication are human decisions, and a person is accountable for what appears on this page. How we work →
Discussion
0 comments