Tencent Ships Full Hy3 Under Apache 2.0: Search Agent Champion That Cedes Code to GLM-5.2
Tencent officially shipped Hy3 on July 6, upgrading from the April 23 preview with strengthened reinforcement learning, higher-quality training data, and a commercially usable Apache 2.0 license. The model is live on Tencent Cloud TokenHub, free on OpenRouter for two weeks, and already integrated into WorkBuddy, Yuanbao, Marvis, and ima.
The headline number from Tencent’s own usage data: daily token consumption on Hy3 preview grew 20x since launch. WorkBuddy users actively selecting Hy3 grew 6x. That is the baseline Hy3 full is inheriting on day one.
Where Hy3 Wins
The agentic search numbers are the case for Hy3. On BrowseComp — the benchmark for multi-hop web research used by Anthropic and OpenAI to grade long-chain web agents — Hy3 posts 84.2, ahead of every open-weight model in Tencent’s published table and competitive with Claude Opus 4.8. DeepSearchQA hits 91.0. On MCP-Atlas, the tool-orchestration benchmark covering 100+ real MCP tool combinations, Hy3 posts 79.1 — leading the open field.
The hallucination story also improved. On internal multi-turn benchmarks, Tencent reports the hallucination rate dropped from 12.5% to 5.4% over the preview. Multi-turn issue rate fell from 17.4% to 7.9%. Long-dialogue MRCR jumped from 42.9% to 75.1%. For agents that need to hold multi-step reasoning across tool calls and long context, the full version is measurably better than the preview.
On Tencent’s internal application suite — 90% task-completion rate. That is the number they are using to sell enterprise developers on the platform.
Where Hy3 Loses
Coding, plainly. VentureBeat’s benchmark appendix, drawn from Tencent’s own comparison table, shows GLM-5.2 ahead across the entire agentic coding suite:
| Benchmark | Hy3 | GLM-5.2 |
|---|---|---|
| SWE-bench Verified | 78.0% | 84.2% |
| SWE-bench Multilingual | 75.8% | 83.0% |
| Terminal-Bench 2.1 | 71.7% | 81.0% |
| DeepSWE | 28.0% | 46.2% |
The DeepSWE gap — 18.2 points — is not marginal. For repository-scale software engineering tasks, GLM-5.2 is the better model. Tencent chose to run its comparison against GLM-5.1, the older model; the newer GLM-5.2 maintains the coding lead throughout.
The Infrastructure Argument
The reason to consider Hy3 anyway is compute. GLM-5.2 is roughly 744B total parameters with ~40B active per token. At FP8, that is 744GB of weights alone — an 8x H200 node as the practical production minimum. Hy3 runs on under 300GB. For an organization self-hosting, the difference is one heavily-specced node versus something it can provision today.
Add Apache 2.0 — no regional exclusions, no field-of-use restrictions — and the enterprise calculus sharpens: GLM-5.2 if coding performance is the only criterion and GPU budget allows. Hy3 everywhere else: search-heavy agent workloads, productivity automation, long-document Q&A, multi-turn tool orchestration, and any deployment where hallucination rates and memory footprint matter more than commit-level code generation.
Pricing
API available on Tencent Cloud TokenHub. Free on OpenRouter for two weeks. Developer fees reduced from preview levels. Tencent has not published the post-promotion per-token price.
Context
The preview launched April 23 with SWE-bench Verified at 74.4%, BrowseComp at 67.1%, Terminal-Bench 54.4%. The full version improves SWE-bench to 78.0% and BrowseComp to 84.2% — a 25-point leap on web research, a 3.6-point improvement on code. The architecture is unchanged: 295B total, 21B active, hybrid fast-and-slow-thinking, 256K context, MoE routing.