Frontline Hotspot
Frontline Hotspot

Tencent's Hy3 Is Free Worldwide via WorkBuddy Until Aug 31: A Window to a Frontier Open-Source Flagship

Tencent's flagship open-source model Hy3 (295B MoE) is free for global users through WorkBuddy until Aug 31 (Pacific Time). Hy3 outscored GLM-5.1 in blind testing and contends with DeepSeek V4 Pro on coding and agentic benchmarks -- a window to use a frontier open-source flagship for free.

Published August 10, 20266 min read
<!-- tencent-hy3-free-workbuddy-hotspot | hotspot | Tencent's Hy3 Is Free Worldwide via WorkBuddy Until Aug 31: A Window to a Frontier Open-Source Flagship -->

In August 2026, Tencent opened its flagship open-source model Hy3 to global users for free via WorkBuddy, its overseas AI office agent, through August 31 (Pacific Time). This is not a lightweight model on promo -- Hy3 is a 295B-parameter MoE open-source flagship that outscored GLM-5.1 in blind testing and goes toe-to-toe with DeepSeek V4 Pro on coding and agentic benchmarks. In other words, it is a rare overlap of "frontier open-source flagship + global free + a full month," and for developers and enterprises wanting to put a top Chinese open model on real workloads, it is a window worth not missing.

A note on sourcing

This piece is assembled from Pandaily's "Tencent Expands International Access to Flagship Hy3 Model," several English-language reports, and Tencent's official information, as of 2026-08-10. Chinese-language coverage was not picked up by search, so English sources take precedence; model capability and leaderboard positions shift in real time, so treat Tencent's official disclosures as authoritative. "Free" here specifically means the WorkBuddy channel is free through 2026-08-31 (Pacific Time) -- it is not permanent, and not every channel is free. For the broader backdrop of China's open-source push, see this site's "China's AI Blitz: How Open Source and Cut-Rate Prices Carved Out a US 'Death Zone'."

1. What Hy3 Is: A 295B MoE Open-Source Flagship

Hy3, fully Tencent Hunyuan Hy3, is Tencent's current flagship open-source model, built on a Mixture-of-Experts (MoE) architecture. Its predecessor was Hy3 Preview; the general-availability release made major improvements across reasoning, coding, tool use, and long context. Starting in February 2026, Tencent rebuilt its pretraining and reinforcement-learning infrastructure, shifting the goal from "chasing leaderboard scores" to "practical, deployable in real business" -- this is the key to understanding Hy3's positioning: it was built to work, not to benchmark.

DimensionSpec
Total parameters295B (MoE)
Active parameters/task~21B ("runs like a 20B")
PredecessorHy3 Preview
Open sourceYes
PositioningPractical, deployable in real business

~21B active parameters means Hy3's per-token inference cost is close to a 20B-class model, not a full 295B -- that is the MoE dividend: big-model capability, small-model bill. It is also one reason a model this size can be offered free to global users at all.

2. Why It Is Worth Using Free: Coding + Agentic Contention, Blind-Test Win Over GLM-5.1

Free alone is not enough; the model has to be strong. Hy3 makes its case on three fronts.

1. Blind-test win over GLM-5.1. In a blind test scored by 270 domain experts, Hy3 earned 2.67/4 versus GLM-5.1's 2.51. The gains concentrated in frontend development and data & storage -- both high-frequency, high-value areas in real work, not niche benchmarks.

2. Coding and agentic contention with DeepSeek V4 Pro. On coding and agentic benchmarks, Hy3 trades blows with DeepSeek V4 Pro. Note it has not toppled GLM-5.2 (the current open-source SOTA), but the gap is closing fast. Do not pick by "who is strongest" alone; pick by where your workload lands.

3. Yao Shunyu and the "second half of AI." Yao Shunyu returned to China and joined Tencent; Hy3 is his first major release there. His "second half of AI" thesis, in one line: sharpen models inside real business and complex scenarios, not just on lab benchmarks. That explains why Hy3's gains show up in tool use, long context, and frontend dev -- the "unglamorous" work -- because that is the work it was trained on.

Hy3 is not a "kills GPT" marketing narrative; it is an open-source flagship forged inside real business -- which is exactly why it is worth using free: the tasks you throw at it may well be the same class it was trained on.

3. How to Use the Free Policy: WorkBuddy, Free Worldwide Through Aug 31

Here is the hook. Hy3 is exposed through three channels:

ChannelWhat it isFree status
WorkBuddyTencent's AI office agent (built on CodeBuddy)Free for global users through 2026-08-31 (Pacific Time)
Tencent Design MioraTencent's overseas design productHy3 accessible
Tencent Cloud TokenHubTencent Cloud's overseas API serviceHy3 accessible

The focus is WorkBuddy. It is a productivity agent desktop workstation that connects GitHub, Gmail, Slack, Notion, and other common tools -- positioned as an "AI office agent" that does not just chat but calls tools and runs task chains. The free offer is part of Tencent's overseas AI push: one month of free flagship access in exchange for developer trial and awareness of Tencent's AI ecosystem.

For developers and individuals wanting to try a Chinese open-source model, the easiest entry point is WorkBuddy: sign up and use it, no need to self-host a 295B model or pay API fees. To pin Hy3 so WorkBuddy does not auto-switch models, see the "turn off auto-select, pin manually" section of this site's "WorkBuddy Field SOP."

4. How It Ranks Against GLM-5.2 / DeepSeek V4 Pro

This is the most practical selection question. A no-hype read:

ModelCurrent positionReason to pick it
GLM-5.2Current open-source SOTAWhen you want the strongest open model, full stop
DeepSeek V4 ProCoding/agentic strengthTrades blows with Hy3 on coding and agent tasks
Hy3Pragmatist, gap closing fastBlind-test lead over GLM-5.1 in frontend/data; and free right now

One line: Hy3 has not unseated GLM-5.2 as SOTA, but the gap is closing fast, and it already holds local edges in specific areas (frontend, data & storage). Add "free through Aug 31" and the value window is real. For a fuller cost-performance comparison, see this site's "LLM Cost-Performance Comparison Review." The core selection logic is not "who is strongest" but "where your workload lands, what your budget is, and how long the window lasts" -- the shift from "capability lead" to "cost structure" is detailed in this site's "China's AI Blitz" piece.

5. Who It Is For, and How

1. Developers. If you want to run a 295B-class open flagship for free -- for prototyping or head-to-head testing -- this is the best window. Sign up for WorkBuddy, pin Hy3, and run your own real tasks (not benchmark puzzles) to see how it actually performs in your scenario. This batch's "Agnes Free Multimodal API Review" and "Zero-Cost Multimodal Workflow SOP" cover the same "free + compose" mindset and are worth reading alongside.

2. Enterprises. If you are evaluating whether a Chinese open-source model can enter production, use this month for a POC. Test three things: tool-use stability, long-context retention, and frontend/data task quality. Hy3's "pragmatist" positioning means it invested most heavily in exactly these three -- a good fit.

3. Individuals curious about Chinese open source. Skip the pain of self-hosting 295B (most people cannot run it); go straight to WorkBuddy's cloud entry and try the flagship at zero cost. After the free window, decide whether to move to TokenHub's paid API or switch models.

6. Three Pitfalls

Pitfall 1: The free window ends Aug 31 -- do not wait. It expires August 31, Pacific Time, not the end of the month in your local zone. If you want to test, test now; do not pile into the last two days when network and quality both degrade. After it ends, Hy3 in WorkBuddy reverts to credit consumption or switches to another model.

Pitfall 2: WorkBuddy's auto-select is unstable -- pin Hy3. WorkBuddy auto-switches models by task and frequently drops you onto a "dumbed-down" one. To reliably use Hy3, pin it manually. Steps are in this site's "WorkBuddy Field SOP."

Pitfall 3: You cannot self-host the full 295B -- use the cloud. Hy3 is a 295B MoE; even at ~21B active, full local deployment demands serious compute. Developers and individuals should not attempt local self-host -- go through WorkBuddy's free cloud entry. Enterprises with GPU clusters can consider self-hosting.

7. FAQ

Q1: Is Hy3 really free? Until when? A: Via the WorkBuddy channel, free for global users through 2026-08-31 (Pacific Time). Note: free applies to the WorkBuddy channel, not all channels permanently; Miora and TokenHub billing follows Tencent's official terms.

Q2: Is Hy3 stronger than GLM-5.2? A: No. GLM-5.2 remains the current open-source SOTA. But Hy3's gap is closing fast, and it already beats GLM-5.1 in blind testing on frontend development and data & storage. Combined with the free window, its value is high.

Q3: Hy3 or DeepSeek V4 Pro? A: The two trade blows on coding and agentic benchmarks -- neither dominates. Run each on your own real tasks and see which is more stable in your scenario. Right now Hy3 is free and DeepSeek V4 Pro is not, so prioritize free Hy3 as your comparison baseline.

Q4: Who is Yao Shunyu, and what is his relation to Hy3? A: Yao Shunyu returned to China and joined Tencent; Hy3 is his first major release there. He proposed the "second half of AI" thesis -- sharpening models inside real business and complex scenarios rather than only on lab benchmarks. Hy3's gains in tool use, long context, and frontend dev are products of that thesis.

Q5: What happens after the free period? A: After Aug 31, Hy3 in WorkBuddy may revert to credit consumption or switch to another model. To keep using Hy3, go through Tencent Cloud TokenHub's paid API; to stay free, watch for Tencent's subsequent promotions or switch to other open-source models. This site will keep tracking.


References

  • Pandaily - Tencent Expands International Access to Flagship Hy3 Model
  • Multiple English reports: Tencent ramps up overseas AI push via Hy3 / WorkBuddy
  • Tencent official information (Hy3 parameters, WorkBuddy positioning, free policy)
  • This site: "WorkBuddy Field SOP" - /en/workbuddy-setup-pitfalls-sop (how to pin HY3 for stability -- read this if you want to use Hy3 free without the auto-switch headaches)
  • This site: "China's AI Blitz: How Open Source and Cut-Rate Prices Carved Out a US 'Death Zone'" - /en/china-ai-blitz-death-zone-hotspot
  • This site: "LLM Cost-Performance Comparison Review" - /en/ai-llm-cost-performance-comparison-review
  • This batch: "Agnes Free Multimodal API Review" - /en/agnes-free-multimodal-api-review
  • This batch: "Zero-Cost Multimodal Workflow SOP" - /en/zero-cost-ai-multimodal-workflow-sop
  • This piece is assembled from public reporting (2026-08-10), not a lab reproduction; model capability and leaderboard positions follow official sources

This article is AI-assisted and human-edited. Last updated: 2026-08-10

FAQ

Is Hy3 really free? Until when?
Free via the WorkBuddy channel for global users until 2026-08-31 (Pacific Time). Note it's the WorkBuddy channel that's free, not all channels permanently; Miora and TokenHub billing per Tencent's official terms.
Is Hy3 stronger than GLM-5.2?
No. GLM-5.2 remains the current SOTA open-source model. But Hy3's gap is closing fast, and it already outscored GLM-5.1 in blind testing on frontend dev and data storage, plus it's free during the window -- great value.
Hy3 vs DeepSeek V4 Pro?
The two trade blows on coding and agentic benchmarks. Run your own real tasks on each to see which is more stable for your scenario. With Hy3's free window now, prioritize it as a comparison baseline.
Who is Yao Shunyu and what's his relation to Hy3?
Yao Shunyu returned to China and joined Tencent; Hy3 is the first major result under his lead. He proposed the "second half of AI" idea -- refining models inside real business and complex scenarios. Hy3's gains in tool use, long context, and frontend dev are products of this.
What happens after the free period ends?
After 8/31, WorkBuddy may switch Hy3 back to credit consumption or another model. To keep using it, go through Tencent Cloud TokenHub paid API; to stay free, watch Tencent's future promos or switch to other open-source models.

Related

Frontline Hotspot

Alibaba Shows Its Qwen4 Hand Early: Qwen3.8-Flash-Next Ships 125B Weights, But No License File

On August 26 Alibaba released Qwen3.8-Flash-Next: a multimodal MoE model that doubles as an early preview of the Qwen4 architecture - the same role Qwen3-Next once played for Qwen3.5. The main model is 125B parameters with an extra 51B of N-gram embeddings, activating just 6B per token; training costs about one ninth of Qwen3.7-Plus while delivering stronger coding and office performance. Four upgrades, unpacked: GDN compresses history while QSA uses a compressed indexer to pick important context at micro-block granularity; Gated Residual widens the residual stream into four branches; the N-gram embedding table can be offloaded to host memory and overlapped with compute via async prefetch; and the optimizer switches to Muon. Native context is 262,144 tokens, extensible to 1M with YaRN. The production Qwen3.8-Flash lists at \$0.16/\$0.47 per million tokens on QwenCloud (sources differ slightly; defer to the official site). The real open question is licensing: the GitHub repo ships no LICENSE file and its license field is None, the README simply points to the Hugging Face or ModelScope model page, and the community is already asking "why isn't it Apache 2.0?" - this article marks it unconfirmed, so verify the model page before any commercial use.

Aug 30, 20266 min read
Frontline Hotspot

Tencent Open-Sources Hy4 Preview: A 770B Flagship That Helped Train Itself

On August 28, Tencent released and open-sourced its new flagship Hy4 preview (770B total / 49B active MoE, 78 layers): Gated DSA sparse attention + IndexCache cross-layer index reuse + iHC identity Hyper-Connections, with the README openly stating the architecture is "inspired by DeepSeek and GLM". A native MTP layer enables 3-token speculative decoding, context spans 1M tokens, and BF16+FP8 weights ship under Apache 2.0. In Tencent's internal blind eval, 163 experts scored 203 engineering tasks at 2.99/4.00, edging out GLM-5.3 (2.92) and Kimi K3 (2.94, both internal-caliber numbers). The headline is the early loop of recursive self-improvement: the model took part in automating optimization of its own training methods, data strategies, eval frameworks and low-level operators, and autonomously lifted inference end-to-end throughput by 31.8%. OpenRouter snapshot pricing: $0.834 input / $2.501 output per million tokens; free for two weeks on WorkBuddy/CodeBuddy.

Aug 29, 20266 min read
Frontline Hotspot

Alibaba's Qwen3.8-Max: 2.4T-param MoE flagship that programs autonomously for days

On 2026-08-03 Alibaba Tongyi released Qwen3.8-Max: a 2.4T-param MoE flagship with 1M context (991K input / 131K output), native vision across plan-execute-verify, positioned to "autonomously program for over ten days delivering complete projects." Pricing: ¥12/M input, ¥36/M output, explicit cache hit ¥1 (1/12 of uncached). Three entry points: blog / Qianwen platform / Qwen Studio.

Aug 3, 20264 min read