Hy4 flash Official Release EvaluationProvider report
2026-11-18Hunyuan 4 flash (Hy4 flash) only: 18 metrics — the release appendix's 8 WorkIndex internal sets, 7 Open Agent Benchmark items and 2 multimodal items, plus the WorkIndex composite taken from the post body and figure. All are vendor-reported results at the highest available reasoning setting; the vendor's per-task cost in CNY is kept in raw.cost_per_task_cny. The WorkIndex sets were run on workbuddy cli and differ from the Hy4 report's PPTGen-Bench and BuddyVQA Atlas setups, so they get their own criteria; the Open Agent Benchmark and multimodal items reuse the Hy4 report's criteria (its Hy4 column matches the scores already on the site row by row), with cost and setup recorded in raw; note that ProgramBench here uses mini-swe-agent, a different harness from the one marked in the Hy4 report. Comparison-model columns (Hy4, Hy4 preview, DeepSeek-V4.1-Flash-1017, Qwen3.8-Flash-Next, GLM-5.3-Flash, GPT-6 Luna) are not imported. List API pricing is ¥0.5 input / ¥0.02 cached input / ¥2 output per million tokens; TokenHub's limited-time 50% discount through November 30, 2026 is not recorded. Weights are open under the Tencent Hy License.