On July 12, Tencent's Hunyuan team officially announced the release of Hy3, a fast/slow-thinking hybrid MoE architecture model with 295 billion total parameters, only 21 billion active parameters, a 256K context window, and full open-source release under Apache 2.0. Hy3 scored 2.67/4 in an internal blind test with 270 experts (GLM-5.1 scored 2.51/4), leading in frontend, data & storage, and CI/CD categories. Hallucination rate dropped from 12.5% to 5.4%, and common-sense error rate fell from 25.4% to 12.7%. Most notably, the agent task success rate on WorkBuddy jumped from 72% to 90%, average completion time shortened by 34%, and the Marvis Agent reached a 93.7% task completion rate.
Hy3 is now integrated into WorkBuddy, CodeBuddy, Yuanbao, Marvis, ima, WeChat Official Account AI avatars, WeChat Reading, and the WeGame game assistant. The API is live on Tencent Cloud TokenHub, with additional overseas API platforms to follow. Pricing: 1 RMB per million input tokens, 4 RMB per million output tokens. On July 8 at 10 AM, Hy3 overwhelmed WorkBuddy's compute capacity with queue rates exceeding 50%; official emergency scaling restored service by the morning of July 9.
1. Architecture: 295B/21B MoE Is Pragmatism, Not Parameter Inflation
Hy3's technical choices look distinctly "counter-consensus" for mid-2026:
| Vendor | Flagship | Approach | |--------|---------|----------| | DeepSeek | V4 series | Very-large-scale MoE + inference optimization | | Qwen | 3.7 | Dense + MoE dual versions | | GLM | 5.2 | Dense model + code specialization | | Meta | Llama 4 | Very-large-scale MoE | | Tencent Hy3 | 295B/21B MoE | MoE + product-driven + Apache open source |
Hy3's core thesis: rather than training a model that is average on every dimension, train one with strong fundamentals plus fine-tuning for specific products. The 295B total parameters provide ample knowledge reserves, 21B active parameters keep inference costs comparable to much smaller models, and feedback loops from 50+ internal products drive continuous tuning.
Key technical design points:
- Fast/slow thinking fusion: switchable fast/slow thinking modes, sharing lineage with Hunyuan-A13B
- 256K context: covers full codebases of medium-sized projects
- Substantially increased post-training: greater post-training compute and data quality versus Hy3 preview
- Apache 2.0 license: free commercial use worldwide; available on HuggingFace and ModelScope
- Task success rate: 72% (preview) → 90% (final)
- Average completion time: shortened by 34%
- Daily token consumption for autonomous selection of Hy3 preview grew 20x
- Users autonomously choosing Hy3 preview on WorkBuddy grew 6x
- Direct delivery of PPT, Word, Excel, PDF, HTML files
- Entirely free (a product strategy that will drive further usage)
- Task completion rate 93.7%, up 12.7 points from preview
- Significant improvements in speed, stability, and cost
- More reliable multi-agent collaboration
- Hy3: 2.67/4 vs GLM-5.1: 2.51/4
- Leading categories: frontend, data & storage, CI/CD
- Hallucination rate: 12.5% → 5.4%
- Common-sense error rate: 25.4% → 12.7%
- WorkBuddy compute saturated, peak queue rate above 50%
- Overnight capacity expansion; service restored July 9 morning
- No large-scale user churn — "Hy3's productivity value was worth the queue"
- Capability line: parameter scale, long context, reasoning depth, benchmark scores
- Engineering line: cost, latency, reliability, debuggability
- Fully open source (Apache 2.0) + near-free pricing (input 1 RMB/million tokens — roughly 1/10 to 1/30 of OpenAI's prices for domestic developers)
- Deep Tencent ecosystem integration (WorkBuddy/CodeBuddy/Yuanbao/Marvis/ima/WeChat/WeChat Reading/WeGame)
- An engineer-driven culture under Yao Shunyu's leadership — a key variable in Tencent AI re-entering the top tier in mid-2026
- Pure vision is a weak point (video understanding, complex chart parsing)
- 256K context is sufficient but trails Claude Code / GPT-5.6 Sol at 1M-window scales
- Tool-calling reliability (structured output, strict mode) needs more third-party validation
- The 72%→90% success rate is Tencent's own data; third-party blind tests are not yet available
- The 50%+ queue rate suggests potential capability degradation under high concurrency
- The sustainability of Yuanbao Agent's free tier is an open question
- Behind the 18-point success-rate gain: which tasks went from 30%→80% (easy) vs 60%→75% (hard)? Distribution matters more than averages
- Is the "34% time reduction" wall-clock or step count? Can it be decomposed into model latency + tool-call latency + network I/O?
- Independent data from HuggingFace / LMSYS in August will be needed
Yao Shunyu (Tencent Hunyuan chief researcher) argues: "In the agent era, sufficient capability + low cost + thorough open-sourcing wins developers." The July 8 WorkBuddy capacity crunch validated this judgment.
2. Real-World Data: Hy3 Optimizes for Engineering, Not Leaderboards
Tencent's official figures come from two months of real production environments:
WorkBuddy agent tasks (automated script generation + workflow orchestration)
Yuanbao Agent (complex tasks in daily conversation)
Marvis Agent (file editing, file management, computer diagnostics)
270-expert internal blind test (real work scenarios)
Capacity crunch (July 8)
3. Analysis: The Agent-Oriented LLM Pivot
Hy3's official positioning is an "Agent-oriented LLM" — a genuine engineering judgment, not marketing.
Two main threads in LLM evolution:
Starting in 2026, these lines diverged. The capability line serves unlocking new abilities; the engineering line serves production usability (95% success rate, 3-second response, $0.01/task, 99.9% SLA). Hy3 invested everything in the engineering line, at the cost of not being SOTA on some capability dimensions (pure vision, complex math proofs).
Compared to contemporaneous frontier models, Hy3's differentiators:
4. Why This Is a Pivotal Moment for AI Coding
Three parallel trends in H2 2026:
1. Chinese open-source models are breaking the price-performance ceiling — DeepSeek V4 Flash RL training on AMD Instinct MI355X (7/9), Meituan LongCat-2.0 open-sourced on a 50,000-GPU domestic cluster (7/9), Xiaomi MiMo-7B coding competition open source 2. Agent evaluation shifting from benchmarks to production data — Hy3 proves itself with real WorkBuddy success rates, part of a broader shift toward real-work-scenario evaluation suites 3. Domestic AI coding toolchain acceleration — Tencent's "four-piece set" (CodeBuddy + Hy3 + Yuanbao Agent + Marvis); Alibaba's Page Agent + Qwen; ByteDance's Doubao + Coze; Meituan's LongCat + Tabbit — every major player is locking in internal workflows with its own model + agent platform
5. Risks and Open Questions
Hy3's limitations:
Credibility of Tencent-side agent metrics:
Is agent evaluation being hijacked by "engineering metrics"?