Tencent officially launched Hy3 on July 6, 2026, releasing a 295-billion-parameter open-source model under an Apache 2.0 license that it claims reaches 90% task resolution on enterprise agentic workflows. That combination of frontier performance, commercial-friendly licensing, and low cost makes Hy3 one of the more consequential model releases of the year for business AI teams weighing build-vs-buy decisions.

What Hy3 Actually Is

Hy3 is a Mixture-of-Experts model, meaning only 21 billion of its 295 billion parameters activate on any given query. That architecture keeps inference costs manageable while giving the model access to specialised capabilities across a broad range of tasks. It supports a 256,000-token context window, which handles most enterprise document sizes, long call transcripts, and multi-turn agentic reasoning chains without truncation.

The model is the full release of a preview Tencent shipped internally in April. In the months between, Tencent ran it through 50 internal product teams and made measurable improvements: hallucinations dropped from 12.5% to 5.4% and commonsense reasoning errors fell from 25.4% to 12.7%. Those are meaningful numbers for enterprise deployments where factual accuracy and consistent reasoning matter more than raw benchmark scores.

How It Benchmarks

On agentic tasks, the numbers are credible. Hy3 scores 84.2 on BrowseComp and 91.0 on DeepSearchQA, competitive with Claude Opus 4.8 and GPT-5.5 on those evaluations. It leads the open-model field on tool orchestration, posting 79.1 on the public MCP-Atlas benchmark set, which is the closest thing the industry has to a standardised test for agent-to-tool calling reliability.

On Tencent’s own WorkBuddy enterprise platform, the model achieves a 90% agent task resolution rate. That number comes from Tencent, so apply appropriate scepticism, but the fact that it is also scoring competitively on third-party benchmarks gives it more credibility than a purely self-reported figure.

Scientific reasoning performance is also strong: 90.4 on GPQA Diamond and 72.0 on USAMO 2026. For data and research teams running analytical workflows, that signals reliable performance on complex multi-step problems.

The Pricing and Licensing Picture

The commercial terms are where Hy3 gets interesting for enterprise teams. On Tencent Cloud, the pricing is set at 1 yuan per million input tokens and 4 yuan per million output tokens, which at current exchange rates puts it significantly below frontier model pricing from OpenAI and Anthropic. On OpenRouter, a free tier exists through July 21, 2026.

Apache 2.0 is the most permissive enterprise-friendly open-source licence available. It allows companies to deploy, modify, and build commercial products on top of Hy3 without royalties, without usage restrictions, and without sharing modifications back. For organisations that need to run models inside their own infrastructure for security, compliance, or data sovereignty reasons, that matters.

The model is already integrated across Tencent’s own product stack, including WeChat, WorkBuddy (their enterprise productivity platform), CodeBuddy, Yuanbao, Marvis, and ima. That real-world deployment at Tencent’s scale provides a different kind of validation than lab benchmarks alone.

What This Means for Business

The open-source frontier is catching up faster than expected. Hy3 adds to a pattern that has been building through 2026: open models from Chinese labs are closing the gap on closed frontier models faster than most enterprise AI roadmaps assumed. Earlier this year, Z.ai’s GLM-5.2 saw 27x growth in token volume on Vercel in its first week. Hy3 is more capable on agentic tasks than GLM-5.2, at roughly half the effective parameter count.

For businesses, this creates a genuine evaluation decision that did not exist six months ago. The question is no longer simply “which frontier API should we use?” It is “should we host our own model, and is it now capable enough to meet our requirements?”

Agentic reliability is the test that matters. The 90% task resolution figure and strong MCP-Atlas scores are the numbers to watch. Most enterprise AI failures happen in agentic settings where models need to sequence tool calls, maintain task state, and recover from errors across multiple steps. Benchmark performance on static question sets predicts far less than performance on these integrated workflows. Hy3’s numbers on agentic evaluations are notably stronger than its overall benchmark position would suggest.

Cost pressure on enterprise AI is real. Several reports from Q2 2026 noted that enterprises were experiencing significant bill shock from agentic AI workloads, where token costs compound across multi-step tasks. A model that runs at a fraction of the cost of frontier APIs, can be self-hosted, and performs competitively on agentic tasks gives finance and IT teams a credible alternative when renewal conversations start.

What it does not solve. Hy3 still needs a team with the infrastructure knowledge to deploy and manage it. A self-hosted model requires GPU capacity, serving infrastructure, monitoring, and ongoing maintenance. For smaller organisations or teams without ML operations experience, the total cost of running Hy3 self-hosted may not be lower than API pricing when compute and engineering time are included. Managed API access through Tencent Cloud provides a middle path, but organisations in markets with restrictions on Chinese cloud infrastructure will need to factor that in.

The Bigger Picture for Enterprise AI Strategy

The release of Hy3 reinforces something Enterprise DNA has been watching closely: the competitive advantage in AI is shifting from model access to model use. When frontier-grade capability is available cheaply or freely, what differentiates results is how well an organisation has structured its data, designed its workflows, and built the operational infrastructure around AI deployment.

Businesses that are still deciding which LLM to bet on are asking the wrong question. The model layer is commoditising. The question that determines actual business outcomes is whether your processes, your data, and your people are ready to direct these tools toward work that generates measurable value.

---

If you want the playbook other teams are using with Claude and Codex right now, grab the free Working With Claude field guide. Download it here.