Hy3
Tencent has unveiled Hy3, their latest open-source large language model, touting competitive performance against significantly larger models and improved agentic capabilities. The model emphasizes reliability, reduced hallucination, and cost-effectiveness, making it a production-grade option. Hacker News is buzzing with comparisons to industry leaders like DeepSeek Flash and GLM, debating its true capabilities, efficiency, and real-world applicability versus its benchmark claims.
The Lowdown
Tencent has officially launched Hy3, a new large language model, following its preview release. The company highlights Hy3's advancements in agentic capabilities, reliability, and efficiency, positioning it as a strong contender against both similar-sized and larger flagship models. This release aims to provide a reliable and cost-effective AI solution for various productivity tasks and product deployments.
- Performance Benchmarks: Hy3 reportedly outperforms similar-sized models and competes with larger open-source flagships, showing significant gains in reasoning, agentic, and long-context tasks, with a specific focus on coding, office work, financial modeling, frontend design, and game development.
- Expert Evaluation: A blind evaluation with 270 experts scored Hy3 at 2.67/4, outperforming GLM-5.1 (2.51/4), particularly in frontend development, data & storage, and CI/CD tasks.
- Reliability Improvements: The model addresses previous issues, improving stability of tool calls, output formats, knowledge grounding, and anti-hallucination (reducing hallucination from 12.5% to 5.4% and commonsense errors from 25.4% to 12.7%).
- Context Retention: Enhanced complex context retention and multi-turn intent tracking, reducing issue rates in internal multi-turn tests from 17.4% to 7.9%.
- Open Source & Cost: Released under Apache 2.0 license on multiple platforms (GitHub, HuggingFace, ModelScope, AtomGit), Hy3 also offers competitive API pricing (1 RMB input, 4 RMB output per 1M tokens) due to hardware-software co-optimizations.
Tencent emphasizes the rapid development cycle, moving from infrastructure rebuild to full release in six months, and commits to continuous improvement in scaling, data quality, and user experience.
The Gossip
Model Matchups & Metrics
Commenters dove deep into Hy3's actual performance compared to a crowded field of LLMs like DeepSeek Flash, GLM-5.2, Mimo v2 Pro, Qwen 3.6 27B, and even various GPT-5.x iterations. While some found Hy3 surprisingly capable for its size and price, others expressed disappointment, suggesting it might be 'benchmaxxed' or underperform in real-world scenarios like security auditing. The 'pelican test' was frequently referenced as a practical benchmark for judging model improvements.
Resource Ruminations & Reality
A significant portion of the discussion centered on the practicalities of running Hy3, particularly its efficiency, quantization resilience, and hardware requirements for local deployment. While the model is lauded for being 'shockingly small for how capable it is,' many noted its substantial memory footprint (requiring 8 GPUs with high memory) still makes it inaccessible for typical consumer hardware. Comparisons to DeepSeek V4 Flash highlighted Hy3's KV Cache inefficiency and the architectural innovations in rival models that enable better quantization and speed.
Product Presentation Puzzles
Beyond technical specs, several users criticized Tencent's handling of the Hy3 launch and user experience. Issues ranged from a lack of clear explanation on the front page, a QR-locked trial chat, and a 'janky' UI demo, to the model's observed drop in OpenRouter rankings. These observations led to questions about its real-world viability and overall accessibility, despite its touted open-source nature and competitive pricing.