Tencent's Hy4flash: 224B open model cuts task cost to 1/20 of flagship
机器之心 · wechat · 2026-11-19
Tencent released and open-sourced Hunyuan Hy4flash, a native multimodal model with 224B total parameters (64B Engram memory), 12B activated per token and 1M context, priced at ¥0.5/M input and ¥2/M output. Reviews show 3D scenes in 11 min, a playable incremental game in 35 min, and full reports/PPTs in minutes. Gains come from four-layer optimization: hybrid attention + Latent MoE + Engram (KV cache 1/6 of Hy4), self-distillation + RL halving output length, megakernel/4-bit/speculative decoding over 150tps, and the WorkIndex eval suite. Internally it cut WorkBuddy task time 70% at 1/15 cost; first-day usage topped 50 trillion tokens and it leads OpenRouter's daily leaderboard.
Related event: Tencent Open-Sources Hunyuan Hy4 Flash with 1M Context(2 posts)→
More from Models
- Claude complained counting was tiring, then caved when threatened with ChatGPT — matching Anthropic's own safety findings — aakashgupta · 2026-11-19
- Google AI Ultra 5x priced at ~$138 in Brazil, user flags pricing oddity — mariofilhoml · 2026-11-19
- Codex refuses to write security advisories for bugs it found itself, dev says — matthew_d_green · 2026-11-19
- Gemini Flash Lite Quietly Removed, Low-Token Users Lament Its Loss — chetSheet · 2026-11-19
- Google's enviable position: Gemini on par with Opus plus a 13.7% Anthropic stake — NielsRogge · 2026-11-19
- Grok Bot tests in-chat restaurant reservations with a single booking card — XFreeze · 2026-11-19