After DeepSeek, another Chinese model maker has made the “Flash tier” its protagonist. On October 23, MiniMax announced the open-sourcing of M3.1 Flash weights: 428B total parameters with 23B active, a comprehensive upgrade over the previous M3. Within 72 hours of launch, real users processed over 100 trillion tokens.
1. Cost down 70%, inference nearly 3x faster
These numbers explain the Flash boom: ① versus the prior M3, M3.1 Flash cuts usage cost by 70%; ② with full-model sparse attention, quantization-aware training and power-aware scheduling, online inference speed nearly triples; ③ on public coding and agent benchmarks, Chinese Flash models now compete with top closed models at far lower prices.
2. Chips and inference stacks join on day one
The release rallied a cross-camp ecosystem: ① on chips, Huawei Ascend, T-Head, Kunlunxin, Moore Threads, NVIDIA and AMD all supported day-one access; ② on inference clouds, Together AI, Fireworks and Nebius joined; ③ vLLM, SGLang, BAAI FlagOS, Hugging Face and ModelScope are pushing adoption across chips and cloud environments.
3. Flash becomes the main battlefield
Two days earlier, DeepSeek updated V4.1 Flash-1017 on October 17, focusing on full-project development. The two camps each lean differently: per public comparisons, DeepSeek V4.1 Flash-1017 scores higher in software engineering benchmarks, while M3.1 Flash stands out in cross-SaaS workflows and cybersecurity, forming a mutually-referencing, continuously competing pair.
Global banks moved fast: Citi said the model consolidates MiniMax's position and may accelerate its B2B commercialization; Jefferies affirmed gains in agent tasks, coding and computer operation with cost efficiency, keeping a Buy; Morgan Stanley called its pricing highly competitive among Chinese Flash models, maintaining Overweight with a HK$900 target. A day earlier MiniMax shares had briefly fallen below the HK$165 IPO price; they rose over 6% after the release.
Key numbers
· 428B total / 23B active parameters
· 100T+ tokens processed by real users in 72h
· Usage cost -70%, inference speed ~3x
· MS target HK$900 (Overweight)
4. Competition moves to delivery efficiency
The Flash boom's essence: model competition is shifting from parameter scale to task delivery efficiency—complete the work while cutting retries, rework and manual takeover. Only parallel gains in capability, speed and cost push tasks into high-frequency use. NineZenith keeps pace with the open-source ecosystem; Zenith-Act supports flexible per-workload model selection in private deployment, helping government and enterprise clients balance cost and capability.
(Sources: ThePaper, Jiemian, People's Finance, IT Home and other public reports)