No. 121
2026-11-17
the AI pulse — brewed, not raw
 AI CRITIQUE
Mavi gökyüzünde dağınık bulutların önünde, mavi cam cepheye sahip ve üzerinde "Tencent" logosu belirgin şekilde yer alan yüksek, modern bir bina.

Tencent launches Hy4 flash, cutting AI task costs by 95% in internal tests

Read the article    English (automatic translation)

Summary

Tencent has introduced Hy4 flash, a 224-billion parameter (12 billion active) MoE-based model designed for daily efficiency tasks. The company claims it reduces per-task costs by 95% compared to the full Hy4 model in internal benchmarks and achieves satisfaction levels close to Claude Haiku 5.5 in a small-scale blind test.

Why it matters

A price of under $0.30 per million output tokens and a 1-million-token context window place significant pressure on mid-tier API pricing in the industry, particularly for high-volume enterprise workloads like document processing and presentation generation. The planned open-weight release and OpenRouter listing could further reduce costs for developers currently reliant on closed mid-tier models.

Highlights

  • The 95% cost reduction was measured only against Tencent's own full Hy4 model and within the company-owned WorkIndex benchmark; no comparisons were made with third-party competitors.
  • The blind test was conducted with only 85 users across 259 tasks, and satisfaction was compared against Claude Haiku 5.5, a mid-tier model rather than a frontier model.
  • With 224 billion total parameters (only 12 billion active per pass via MoE) and a 1-million-token context window, the price of $0.07 per million input tokens offers an aggressive price-performance ratio for this segment.

Critical take

The headline '95% cost reduction' is a self-referential comparison made in an internal benchmark against Tencent's own flagship model. The user satisfaction claim relies on a sample size of 85 users and 259 tasks, which is too small to generalize. Until independent, multi-vendor benchmarks are published, the true competitive advantage remains unverified.

This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.
Back to the issue