Hy3 Local AI vs Cloud REVIEW - The NEW Open Source Qwen 3.7? 🤯

Hy3 Local AI vs Cloud REVIEW - The NEW Open Source Qwen 3.7? 🤯

by xCreate · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 xCreate 👥 26K 📅 July 9, 2026 ⏱ 22 min 👁 7K 📄 expert opinion 🧭 2026-09-16
Available in: English (current) Français

Keywords

Hy3TencentMixture of ExpertsvLLMOpenRouter

Summary

In this review, the host explores Tencent’s Hy3, a 295B-parameter Mixture-of-Experts model released under the Apache 2.0 license. The video opens by presenting benchmark scores that show Hy3 outperforming DeepSeek V4 Pro and GLM 5.2 in math and long-context reasoning, and topping open-source agentic performance on Claw eval. The host then runs a series of practical tests, generating games like Flappy Bird, Minecraft clones, a 3D Earth, an Outrun-style racer, and a Super Mario-like platformer, both locally (at 9-bit quantization) and via OpenRouter. A key focus is the scaling of thinking modes: disabled, low, and high, which produce increasingly elaborate outputs. The review also demonstrates Hy3’s ability to handle image-related tasks by leveraging OCR and classification metadata, since the model is text-only. Math problems and logic riddles are tested, with thinking mode enabling correct solutions. The host highlights the MTP (Multi-Token Prediction) layer as an experimental feature. The video concludes by positioning Hy3 as a strong open-weight alternative to Qwen 3.7, given its permissive license and competitive performance.

168 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides valuable hands-on data on Hy3’s real-world performance, including token generation speeds, quantization effects, and the impact of thinking levels. The host’s testing methodology is transparent, with clear examples of generated outputs and code. Arguments are mostly based on personal observation rather than controlled experiments, but the breadth of tests across domains gives a good qualitative picture. The reasoning about the model’s strengths (high math performance, efficient minimal generation when thinking is disabled) and weaknesses (basic UI in generated apps) is coherent and supported by evidence shown on screen.

Scientific Rigor, Source Quality, Title Accuracy

The video cites primary sources such as the Hugging Face model page, OpenRouter platform, and companion videos, and the description includes these links. The host refers to benchmarks from external leaderboards but does not always specify the source. The title accurately reflects content, and the review is internally consistent. No contradictory sources are mentioned. The presentation is scientifically casual but not misleading, with clear disclaimers about the model’s text-only nature and the makeshift image classification workaround.

182 words

Title / Content Match

The title accurately reflects the content: the video compares local and cloud versions of Hy3 and positions it against Qwen 3.7, which is addressed through benchmarks and direct tests.

Quality & Reliability

7/10

The review is based on hands-on testing across multiple benchmarks and practical coding tasks, with transparent reporting of token counts and speeds. However, it is largely anecdotal and lacks a rigorous statistical methodology, and the host's enthusiasm may introduce bias.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

External References

Contribution & Novelties

This review offers a practical, code-focused evaluation of a newly released open-weight model, contrasting local and cloud inference. It contributes insights into the model’s behavior across thinking modes and its creative coding abilities, which are often more illustrative than pure benchmark numbers.

Pour aller plus loin :

  • Mixture of experts — Explains the architecture behind Hy3, which is key to understanding its performance.
  • Apache License 2.0 — The permissive license enabling commercial use, a major factor for adoption.
  • vLLM — A high-throughput inference engine mentioned implicitly for running large models locally.
  • Ollama — Popular local LLM runner, likely relevant for users wanting to try Hy3 locally.

106 words

Radar Profile

The radar shows a balanced profile with high scores in information quantity and technical level, reflecting the detailed hands-on testing. Quality and reliability are slightly lower due to the anecdotal nature of the review.

Reliability 6/10