Tencent Hy4
Tencent’s new Hy4 open-weight language model is being evaluated against rivals like DeepSeek, Kimi, and OpenAI/Anthropic models on both price and capability, with many noting it lands near the cost–accuracy frontier but can be slow and overly thorough. Commenters highlight its strengths in data analytics and game development, while questioning its cost competitiveness per task and the impact of safety-driven training choices that exclude cyber‑offense and other “taboo” domains. The release also feeds into broader debates about Chinese AI progress, export controls, surveillance norms, and whether cheaper open models will erode the advantage of Western proprietary systems.
Model performance & benchmarks
- Hy4 is generally viewed as strong among open-weight models but still behind top proprietary models on difficult math, long-horizon coding, self-improvement, and advanced vision.
- Compared to Kimi 3.1 Pro, Hy4 is said to be slightly weaker but significantly cheaper and trained with fewer tokens (770B vs 2.8T), with some arguing larger models mainly improve “taste” and long-context behavior.
- On specific benchmarks:
- FrontierCode 2.0 score of 31 is seen as roughly comparable to GPT‑6.1 Sol and Opus 5.5 Medium.
- Plotly’s data analytics benchmark places Hy4 on the cost/accuracy frontier between GPT‑6.1 Sol and Terra but notes it is very thorough and very slow, often issuing many queries.
- On PacBench (Pac-Man game agent), Hy4 ties DeepSeek V4.1 Flash in game score while being about 7× cheaper but somewhat slower.
Pricing, efficiency & cost per task
- Some argue Hy4 is “not cost competitive” vs DeepSeek based on posted token prices; others say per-token pricing is misleading and cost per task is what matters.
- Experience with DeepSeek V4.1 Flash indicates it can be cheaper in practice due to very low cached-token prices and high cache hit rates.
- A linked “intelligence-efficiency” analysis is criticized for running DeepSeek on “Max” instead of “High” and for using benchmarks that don’t reflect multi-turn, large-context software engineering work.
Safety, offensive capabilities & content filtering
- Hy4’s training intentionally filters out offensive content such as exploit development, reverse engineering, and penetration testing, disappointing users who rely on such capabilities.
- Some contrast this with approaches that train on cyber offense and then try to bolt on safety; others praise Tencent’s choice as a more straightforward mitigation.
- There is discussion of Chinese export controls (“Announcement 48”), with uncertainty about how they will affect open-weight models, especially for biomedical applications.
Open vs proprietary models & ecosystem trends
- Several comments view Hy4 and other recent open models (DeepSeek, Kimi, Qwen, Particle‑1, Ember‑1) as pushing open weights back onto the cost/accuracy frontier.
- Post-training on existing strong models is noted as increasingly effective.
- Some predict that open Chinese models will surpass US labs within a year; others insist US frontier models (Opus/Sol/Astra) remain clearly ahead at the high end and that cheaper models actually increase demand for top-tier inference (Jevons paradox), a point others dispute as not universally applicable.
Game development focus
- Hy4 is heavily tuned for game dev: feature implementation, gameplay logic, state tracking, asset pipelines, and support for Unreal, Unity, and Godot, including multiplayer.
- This is seen as a natural fit given Tencent’s large global game portfolio and stakes in major studios.
Geopolitics, economics & future hardware
- Some argue the US economy’s outlook is tied to whether OpenAI/Anthropic stay ahead of Chinese models; others find this exaggerated or “doom-y.”
- There are analogies to 2008 “too big to fail” dynamics and to Silicon Graphics: if hardware, quantization, and architectures keep improving, near‑SOTA LLMs may eventually run cheaply on consumer machines, pressuring current SaaS-style business models.
- Others counter that many firms can build capable hardware and models worldwide, so AI benefits are not locked to any single country or lab.
Surveillance, technocracy & attitudes to technology
- Hy4’s vision demos (e.g., satellite imagery, geolocation) prompt comments that Chinese releases seem less concerned about looking like dual-use or military tech.
- Several anecdotes contrast China’s visible high-tech infrastructure and pervasive surveillance with Western concerns about privacy and rights; views diverge:
- Some describe China as a kind of technocracy where people broadly perceive technology as improving daily life.
- Others emphasize censorship, monitoring (e.g., phone telemetry scandals), and the risk that intrusive systems, once deployed, are hard to roll back.
- Polls on perceived democracy and governance quality are cited to argue that Chinese citizens often feel better represented than Western publics; others remain skeptical, noting propaganda, uneven development, and human-rights concerns.
Usability, naming, and trust
- The official Hy4 site is criticized as clunky (e.g., hard-to-read/zoom benchmarks).
- There are jokes about more colorful Chinese model names and references to “big brother”-style branding.
- Some express distrust about sending data to Chinese providers; others push back that such shallow dismissals add little without concrete arguments.