Cloud-led Future · Intelligence-driven Hotline +86 21 6228 0217 中文 | EN

Alibaba Qwen4-Max and Tencent Hunyuan Hy4 Flash Launch; Cost per Task Becomes New Battleground

Overview

On November 16, Chinese large-model releases clustered together: Alibaba unveiled its next-generation foundation model Qwen4-Max, while Tencent’s Hunyuan launched Hy4 Flash. Both put “cost per task” front and center.

Qwen4-Max has 2.4 trillion total parameters using a new sparse-attention MoE architecture, activating about 71 billion parameters per token, roughly 25% fewer than its predecessor. It features native vision and a one-million-token context window. The company reported a sixth-place finish on the CodeArena front-end blind test, cache-hit pricing as low as 0.5 yuan per million tokens, and average cost per task down about 60% from the prior generation.

Tencent’s Hy4 Flash uses a 224B-total, 12B-active MoE architecture with 1M context, priced at 0.5 yuan per million input tokens, 2 yuan output, and 0.02 yuan on cache hits. The company said blended cost per task is about one-twentieth of its flagship Hy4, with decoding above 150 tokens per second, and plans to open-source it later.

The same day, DeepSeek cut free-tier usage limits citing compute constraints, as domestic models advanced on both open-source and cost-reduction fronts.

Background & Interpretation

1. Technical context and evolution

Competition is shifting from “parameters and leaderboards” to “how much work a million tokens can do.” The technical basis is sparsity and cache management: enlarging total parameters while shrinking active ones preserves capability while cutting per-inference cost, and optimizing KV cache and hit rates reduces overhead in long-context scenarios. Qwen4-Max trimming active parameters and Hy4 Flash sharply compressing KV cache follow the same efficiency path.

2. Core drivers and mechanics

Demand-side momentum comes from AI agents — the most token-hungry format, where long sessions and multi-round tool calls greatly amplify costs. Industry disclosures put some models’ weekly call volume in the tens of trillions of tokens, making cost sensitivity acute. As a result, cache-hit rate, tokens per task and end-to-end latency are replacing raw benchmark scores as selection criteria. Building evaluation sets on in-house product tasks also reflects declining correlation between public leaderboards and real-world tasks.

3. Market and supply-chain effects

Two vendors cutting per-task prices on the same day via different routes shows price competition has moved from “price per million tokens” to “price per task.” Upstream, this implies structural change in inference-compute demand: cheaper calls expand total volume but compress the compute premium per call. Smaller model vendors face heavier pressure, while leaders with in-house architectures and engineering optimization can win enterprise migration on price-performance.

Implications & Outlook

Near term, API price cuts should accelerate adoption and call-volume growth for agent applications. Over the medium term, whether capability and cost leadership can coexist depends on architectural efficiency rather than compute stacking. Watch real enterprise migration volumes, the actual cost impact of cache-hit rates, and the release cadence of open-source mid-size models.

Takeaways for Industry Participants

AI application teams should use “cost per task and end-to-end duration” as primary selection metrics and treat cache-hit rate as a first-order target for prompt optimization. For model vendors, architectural efficiency and engineering optimization are the keys to surviving a price war. Developers evaluating domestic models should run internal blind tests tailored to their own task distribution rather than relying solely on public leaderboards.

This column compiles industry information and shares technical perspectives; it does not constitute investment advice.