Tencent touts Hy4 as top open model — but weights aren't out and rivals' scores are its own runs
mawcopolow · reddit · 2026-11-04
Tencent claims Hy4 (final version released November 4, 2026) ranks among the best open models, but the weights haven't been released. Its own comparison table shows Hy4 scoring 78.9 on OSWorld 2.1 computer use (near Claude Opus 5.5's 81.8 and Kimi K3.1 Pro's 79.6), 44.4 on Terminal-Bench 4.0 (well below Claude's 66.4), 78.6 on MMMU-Pro, and just 13.0 on its internal Hy-FrontierMath T5 test versus GPT-6.1 Sol's 97.5. Most competitor scores were run by Tencent itself, raising verifiability concerns.
More from Models
- Users complain Claude wastes tokens on endless apologies instead of just fixing errors — RachelVT42 · 2026-11-05
- Hy-4 vs DeepSeek Flash benchmark: quarter the tokens, a third of requests — ShunyuYao12 · 2026-11-05
- The Efficiency Landscape: GPT-6.1 Sol tops value among top 30 AI models, Opus 5.5 priciest at $10.04 — daniel_mac8 · 2026-11-05
- MiniMax teases upcoming M3 Pro model at Seoul meetup co-hosted with ollama — ollama · 2026-11-05
- Why this dev prefers a 27B dense model with 1M context over a big MoE at 128k — Express_Quail_1493 · 2026-11-05
- Wes Roth: OpenAI covert underperformance report, weight-space programming, Qwen critique — Wes Roth · 2026-11-05