Contextual Placement: Tencent’s Infrastructural Restart and Product-Driven Design
Tencent announced a fundamental rebuild of its pre-training and reinforcement learning infrastructure in 2026 – a rare signal for organizational change in an established AI lab. This decision was made in January 2026. Forbes reports that the team redesigned this infrastructure with an unusual premise: models should no longer be perfected through isolated benchmark optimization but should be trained directly in product loops. Hy3 is the first model to emerge from this rebuild.
The process was quick. From the infrastructure reset in January to the Hy3 Preview in April and the official release in July 2026 – under 6 months for the entire iteration program. This is remarkable because it suggests that vertical integration of research and product development can accelerate development cycles. Tencent integrated Hy3 directly into product teams: WorkBuddy (productivity agent), Yuanbao (AI assistant), CodeBuddy, and ima (knowledge management). These products generated real-time feedback that flowed directly into the training.
Architecture: Mixture-of-Experts for Cost Efficiency Without Performance Loss
Hy3 uses a mixture-of-experts architecture with a total of 295 billion parameters, of which only 21 billion are active during each inference. According to Tencent, an additional 3.8 billion parameter layer is added for multi-token prediction (MTP) – a technique for faster token generation through speculative decoding.
The model manages a routing system over 192 specialized experts, with the 8 most competent activated per token. This architecture follows a design principle that is discussed in Forbes: the 300B parameter limit is intentional. Tencent found that beyond one trillion parameters, latency suffers more from multi-node deployment than marginal capacity gains justify.
Hy3 supports a 256K token context window – sufficient for long document analyses, code refactoring across multiple files, and complex multi-turn workflows.
Performance Metrics: Benchmark Comparisons and Blind Expert Evaluation
On standardized benchmarks, Hy3 is competitive. WinBuzzer reports that the model achieved 74.4% on SWE-bench Verified (Claude Opus 4.6: 80.8%, GPT-5.4: 78.6%) and 54.4% on Terminal-Bench 2.0. On ClawEval – an agent-focused benchmark – Hy3 scored 68.5, surpassing DeepSeek V4 Pro (62.4) and Qwen 3.7 Max (65.2).
However, Tencent emphasizes an alternative evaluation model. The company conducted a blind evaluation with 270 experts who tested Hy3 against GLM-5.1 on real workflows. VentureBeat reports that Hy3 scored 2.67 out of 4 points over 312 valid comparisons, while GLM-5.1 scored 2.51. The strongest advantages were in frontend development, CI/CD, and data & storage.
The company justifies this alternative by stating that public benchmarks test under idealized conditions – complete queries, coherent context. Real users submit fragmented, ambiguous inputs. For a productivity AI, this is the lower bound that matters.
Hallucinations and Reliability: Halved Error Rates Under Controlled Conditions
According to Tencent, the hallucination rate dropped from 12.5% to 5.4% in internal tests. At the same time, errors in common-sense reasoning fell from 25.4% to 12.7%. Pandaily additionally reports that ima – Tencent’s knowledge database agent – achieved system stability of 95.1% in evaluations, with significantly reduced blind repetitions and unnecessary operations.
Important Note: These figures come from Tencent measurements in controlled scenarios. Independent verifications do not exist. The measurements follow the design principle on which Hy3 was built: prioritizing resilience against incomplete or contradictory inputs rather than maximizing benchmark scores.
Product Integration and Real Metrics
Product integration was immediate. Hy3 was introduced into WorkBuddy (Tencent’s main productivity AI in China), CodeBuddy, Yuanbao, and ima. Tencent reports that users of WorkBuddy who actively selected Hy3 Preview increased sixfold since launch. Average daily token usage increased by a factor of 20.
The metrics show where Hy3 demonstrates strength:
- WorkBuddy: Automated script generation and workflow orchestration. Task resolution improved from 72% (with Hy3 Preview) to 90% (with official release). Average execution time reduced by 34%.
- Financial Modeling: Hy3 independently constructed a consolidated cash flow model across 3 regions, 6 reserve blocks, and 5,220 linked cells without hardcoding – each number was a live formula that recalculated with assumption changes.
- Data Modeling: Given 101 SKUs sales data, Hy3 produced Excel modeling with 12 analysis tables and a 30-page presentation with 20 charts.
Open Source Availability and Licensing Implications
A significant difference from the preview: Tencent released Hy3 under Apache 2.0, not under a Tencent-specific community license. On X, this was highlighted by researchers as the real headline news – it signals that Tencent is actively competing for the open AI community.
The model is available on:
An FP8-quantized version reduces memory footprint for edge deployment. OpenRouter provides access with two weeks free since launch time. Other platforms like OpenClaw, Kilo, and Cline are expected to follow progressively.
Implications for the AI Model Market
Hy3 illustrates a trend in the AI industry: efficiency beats scale when training is directly linked to product feedback. Tencent integrated model development with product telemetry – not afterward, but during. This contrasts with the older paradigm: train a large model on general benchmarks, then optimize for use cases.
This has economic implications. A model with 21B active parameters is cheaper to serve than a similarly sized dense model and achieves – according to Tencent – performance of models with 2-5x the parameters. For companies interested in cost efficiency and for the public AI community looking to deploy open-source models in production, this is significant.
At the same time, the emphasis on blind expert evaluation and real product metrics over leaderboards underscores a more mature engagement with AI evaluation. Benchmarks have value, but they do not measure whether a model is reliable under conditions of uncertainty and ambiguity – which is what matters in production.
