Hy Flash
by Tencent · Model · Language & code models
Very cheap and good enough for everyday work, but not frontier-class
Tencent released weights for Hy4 Flash: 224B multimodal MoE, 12B active per token, 1M context, ~64B n-gram embedding table; instruct, FP8 and base checkpoints
Why we say so
Fahd Mirza found it handled files, slides, invoices and multilingual pages well and judged it good enough for everyday work at almost no cost, not frontier-class. Plotly's bench had it beating GPT-6 Luna on cheap data analytics, while HN users are split against DeepSeek V4.1 Flash: one used it for days with good results, another found it no cheaper per task and worse at translation. Tencent's own comparisons show it trailing Hy4 and DeepSeek on terminal coding and complex spreadsheets.
Evidence
- Nov 17Tencent's cost-efficient Hy4 variant: 24B total, 12B active parameters, natively multimodal, 1M-token context, cheaper API than Hy4
- Nov 17Built a single-file multilingual cafe page with two tabs in about 3 minutes; scripts including Chinese, Hindi, Arabic rendered properly
- Nov 17Benchmarks competitive with GLM 5.3 Flash at a much lower price, and a smaller model (224B vs 320B)
- Nov 17Beat GPT-6 Luna on cheap-and-accurate data analytics; answered 43 questions for $0.13
- Nov 17Does not beat DeepSeek V4.1 Flash on cost per task; DeepSeek has very cheap providers and cache hits
Show all 8
- Nov 16Blind user test: near-level user satisfaction with Claude Haiku 5.5, fastest task time, at a small fraction of the cost
- Nov 16In the blind test DeepSeek scored slightly higher on satisfaction but was much slower and cost more
- Nov 16Stays close to pricier models on office work like slides and documents, but trails Hy4 and DeepSeek on terminal coding and complex spreadsheets