The AI large-model race has found a new battlefield: Flash models. On October 17, DeepSeek updated its Flash model to V4.1-Flash-1017, with its web, app and API synchronized to the new model. Two days later, on October 19, MiniMax released the next-generation Flash model M3.1 Flash and previewed a flagship Pro version, China Fund News reported. Both are head model vendors, and their successive Flash updates signal intensifying competition on complex tasks.
Flash models are lightweight, high-cost-performance versions designed to accomplish complex tasks with minimal compute. M3.1 Flash greatly improves capabilities in complex Agent tasks, computer use, software development and cybersecurity. In BankerToolBench, an open benchmark for investment-banking end-to-end workflows, it scored 76.2; its online inference speed improved nearly threefold using a more aggressive sparse-attention architecture, with everyday output above 200 tokens per second and a stable median of 170+.
MiniMax also cut API prices: input 1 yuan and output 4 yuan per million tokens, cache-read 0.1 yuan — more than halving input/output prices versus M3 and cutting cache-read by over 70%. On the DeepSWE v1.1 coding benchmark M3.1 Flash scored 73.7%, entering the industry's first tier.
The two vendors' 72-hour sprint shows domestic Flash models moving from "flagship substitutes" toward becoming the main force in complex tasks — breaking the old assumption that strong coding must rely on flagship models and striking a commercially valuable balance of performance, latency and cost.
Source: China Fund News / 36Kr
