Alibaba and Tencent Release New AI Models on the Same Day, Slashing Per-Task Costs to as Low as 1/8 of Competitors
nashnova research
On November 16 Alibaba released Qwen4-Max and Tencent launched Hy4 flash, both built around drastically lower inference costs — timed two days before their Q3 earnings on November 18-19, putting China's AI pricing war under immediate financial scrutiny.
How much cheaper is Qwen4-Max, exactly?
According to Alibaba's own Work360 benchmark, per-task cost fell from ¥26.58 to ¥11.16 — a 58% drop.
The driver is token efficiency: at high inference intensity, token consumption dropped 24%-35% versus the prior generation; on Terminal-Bench 4.0 it dropped ~48%-49%.
This means → the model isn't just cheaper on paper — it burns less "computational fuel" per task. The savings are real resources, not just billing discounts.
Cached-hit API pricing is down to ¥0.5 per million tokens.
What changed in the 2.4-trillion-parameter architecture?
Qwen4-Max has 2.4 trillion total parameters but activates only 71B per token — in plain terms = the model is huge, but only a small fraction fires on each step, saving compute while preserving capability.
The architecture upgrades its sparse MoE — a design where the model activates experts on demand rather than running all parameters at once — across three axes: attention, residuals, and embeddings.
N-gram Embedding — a technique that expands capacity using short token fragments — adds 328B parameters at relatively low extra compute cost.
Alibaba says Qwen4-Max-Base achieved the lowest bits-per-byte across all five tested models in knowledge, code, Q&A, and STEM benchmarks, with per-token activated parameters down roughly 25%.
How does Tencent's Hy4 flash get costs down to 1/8 of rivals?
Hy4 flash has 224B total parameters but activates only 12B — this means → it is far lighter than Qwen4-Max, positioned squarely on the cost-efficiency frontier.
Tencent's WorkIndex benchmark shows per-task cost at 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash (max tier, peak pricing).
In plain terms = for the same everyday work task, Hy4 flash costs a fraction to a tenth of leading alternatives.
Three technical levers: pre-training shrinks KV cache — the model's short-term memory footprint — to 1/24 of Hy3; post-training uses self-distillation plus reinforcement learning to halve output length; inference-time optimizations push decoding past 150 tokens per second.
Which model targets which use case?
Qwen4-Max targets complex enterprise scenarios: Alibaba cites automotive manufacturing, IPO legal due diligence, and financial research — in the auto case, programming time per workpiece dropped from roughly 475 minutes to 24 minutes, a ~95% reduction.
Hy4 flash targets daily office productivity: 85 internal Tencent users blind-tested it across 27 categories and 259 everyday tasks — satisfaction approached Claude Haiku 5.5, while per-task cost and completion time were the lowest among all tested models.
This means → despite launching on the same day with the same "cut costs" narrative, the two are aiming at different market tiers — one goes after high-end project budgets, the other after everyday desk-level spending.
Why does the benchmarking method matter on its own?
Both companies moved away from public leaderboards, each building internal benchmarks modeled on real work tasks.
Alibaba's Work360 covers 26 primary domains, 182 sub-domains, and 721 fine-grained categories, attempting systematic coverage of industrial and professional scenarios.
Tencent's WorkIndex focuses on output quality for daily work, single-turn output length, and end-to-end task time.
This reflects a broader shift: the industry is moving from "score races" to "real-job outcomes" — whether a model is good enough is increasingly defined by what it costs and delivers on actual tasks.
Launched right before earnings — what should investors watch?
Alibaba reports Q3 earnings on November 18; Tencent on November 19 — both models debuted on November 16, precisely two days ahead.
On API pricing, Alibaba's cached-hit rate sits at ¥0.5 per million tokens; Tencent's at ¥0.02 per million tokens. Hy4 flash also offers a limited-time 50% discount, with open-sourcing planned.
This means → next week's earnings releases aren't just about revenue growth — the key question is whether the cost of waging this AI pricing war shows up on the margin line. The money spent to undercut rivals has to come from somewhere.
市场有风险,内容仅供研究参考,不构成投资建议。
