Key Takeaways:
Minimax (00100.HK) fell 5.92% to HKD301.8 and Z.AI (02513.HK) dropped 7.01% to HKD849.5 on Wednesday, after DeepSeek cut prices across its V4 Flash model and Morgan Stanley warned that competition in the lightweight tier of Chinese large language models will intensify.
Morgan Stanley analysts Gary Yu, Lydia Lin and Yang Liu wrote that the reductions were driven by Alibaba's Qwen3.8-flash and Z.AI's GLM5.3-flash, both launched in August, and by pricing strategy ahead of DeepSeek's own V4.1 Flash release. "Although price competition for Flash models may intensify, gross margin remains crucial," the analysts said, adding that they do not expect a disorderly price war.
The cuts, announced on September 9 and effective the following day, lowered DeepSeek's non-cached input price by 33% to RMB1.5 per million tokens, cached input by 60% to RMB0.03, and output by 11% to RMB6. Even after the adjustment, V4 Flash remains the most expensive of the three mainstream lightweight models. Qwen3.8-flash charges RMB1 for non-cached input and RMB3 for output; GLM5.3-flash charges RMB0.4 and RMB1.4 — less than a quarter of DeepSeek's post-cut output price. DeepSeek's revised rates also sit above its own April launch pricing of RMB1 input and RMB2 output, making the move a partial correction rather than a return to entry levels.
The three models score within five points of each other on the AA Intelligence Index — GLM5.3-flash at 57, Qwen3.8-flash at 56 and DeepSeek V4 Flash at 52 — leaving DeepSeek without a performance premium to defend its higher quote. Parameter counts diverge more widely: GLM5.3-flash carries 320 billion parameters, DeepSeek V4 Flash 284 billion and Qwen3.8-flash 125 billion.
Selling pressure in the two Hong Kong-listed developers was concentrated in short activity. Short selling in Minimax reached HKD472.35 million, equal to 10.08% of turnover, while Z.AI short sales totaled HKD396.97 million, or 12.70% of turnover, according to data as at 16:25 HKT on September 9. Minimax opened 1.56% lower and touched an intraday trough of HKD298.2 on volume of 5.6447 million shares and turnover of HKD1.728 billion. Z.AI opened 0.6% down, hit HKD844 at its weakest and traded 2.1938 million shares worth HKD1.896 billion.
The declines extended a run of analyst caution on Z.AI's revenue trajectory. Jefferies cut its target price on the stock to HKD1,183.79, saying the company's annual recurring revenue guidance beat expectations but questioning whether that pace is sustainable. The trimmed target still implies roughly 39% upside from Wednesday's close of HKD849.5.
DeepSeek's own roadmap adds a second layer of pressure. V4.1 Flash, available for testing through the DeepSeek API until September 10, processes up to 427 tokens per second and handles 3D modeling, game development and spatial reasoning tasks, according to Geeky Gadgets' review of the test build. The model keeps V4 Flash pricing during the trial period, so the near-term competitive threat to peers is capability rather than cost. Reviewers flagged overthinking tendencies and weaker performance than GPT-6 Astra and Fable 5.1 in physics-based simulations.
Morgan Stanley framed the broader shift as structural: the pairing of an expensive, high-quality Pro model with a cheap, fast Flash model is becoming the standard product strategy for LLM vendors, with the Flash layer absorbing the price competition. That leaves gross margin as the binding constraint on how far quotes can fall, since inference costs scale with usage rather than with the price charged per token.
For holders of Minimax and Z.AI, the session's message is that Flash-tier pricing is now a sector-wide variable rather than a company-specific one, and that short sellers are positioned for further downside. The next test comes when DeepSeek publishes commercial pricing for V4.1 Flash after the September 10 testing window closes, and when Z.AI next updates the market on ARR.
This article is for informational purposes only and does not constitute investment advice.
