Alibaba and Tencent Unveil New Models on the Same Day: Qwen4-Max Cuts Per-Task Cost by 58%, Hy4 flash Drops to One-Eighth of Rival's Price

Deep News15:04

On November 16, Alibaba and Tencent each launched new models on the same day. Alibaba released Qwen4-Max, the first flagship of the Qwen4 series, while Tencent Hunyuan introduced the lightweight model Hy4 flash.

According to Alibaba's officially published Work360 benchmark data, the average cost per task for Qwen4-Max fell from 26.58 yuan to 11.16 yuan, a decline of 58%. Tencent said that Hy4 flash's per-task cost on the WorkIndex benchmark is only 1/20 that of Hy4, and just 1/8 that of DeepSeek-V4.1-Flash-1017 (max tier, peak-hour pricing).

Notably, the Q3 earnings reports of Alibaba and Tencent are set to be released this week, on November 18 and November 19, respectively.

Qwen4-Max: 2.4 Trillion Parameters, Task Cost Cut in Half

In an official blog post, Alibaba explained that Qwen4-Max adopts an all-new architecture in the Qwen4 series, with total parameters of 2.4 trillion (2.4T), 71B activated parameters per token, and an additional 328B N-gram Embedding parameters, supporting native vision capabilities and a one-million-token context window.

Behind the cost reduction is an improvement in token efficiency. According to Alibaba's official data, on the Work360 benchmark, under the Xhigh and High reasoning-intensity tiers, Qwen4-Max's token consumption dropped by roughly 24% and 35%, respectively, compared with the previous-generation Qwen3.8-Max; on Terminal-Bench 4.0, the two tiers fell by about 48% and 49%, respectively. Meanwhile, API pricing was lowered across the board, with the cache-hit price seeing a particularly notable reduction, falling to as low as 0.5 yuan per million tokens.

On performance, Alibaba officially stated that compared with the previous-generation Qwen3.8-Max, Qwen4-Max's score on the comprehensive Work360 benchmark rose from 47.74 points to 52.69 points, with clear gains also on several specialized Agent benchmarks such as DeepSWE, AA-Briefcase and HLAB-Vals.

At the architecture level, Qwen4-Max upgraded its sparse MoE model around three directions: Attention, Residual and Embedding. GDN+Qwen Sparse Attention (QSA) reduces the computational overhead of long-sequence attention; Gated Residual (GR) expands the residual stream into four parallel branches to enhance cross-layer information transfer; and N-gram Embedding expands model capacity at relatively low additional computational cost.

Alibaba officially said that across four categories of internal evaluations — knowledge, code, question answering and STEM — Qwen4-Max-Base recorded the lowest bits-per-byte (BPB) among the five participating models, with BPB in each domain falling by roughly 5% to 13% compared with Qwen3.8-Max-Base, while the activated parameters per token dropped by about 25%.

Hy4 flash: Only 12B Activated Parameters, Cost Pressed to One-Eighth of Rivals and One-Twentieth of Hy4

Tencent Hunyuan's Hy4 flash, released on the same day, is positioned as more lightweight. According to Tencent's official introduction, the model has 224B total parameters (including 64B Engram parameters), only 12B activated parameters, and supports multimodality and a 1M context length.

Tencent positions Hy4 flash as "the most cost-effective Hy model to date." On API pricing, input costs 0.5 yuan per million tokens and output costs 2 yuan, and only 0.02 yuan when the cache is hit. On the WorkIndex benchmark, the per-task cost is just 1/13 that of Hy4 preview, 1/20 that of Hy4, and 1/8 that of DeepSeek-V4.1-Flash-1017 (max tier, peak-hour pricing).

To achieve this cost level, Tencent carried out targeted optimizations across multiple stages. In the pre-training stage, it introduced architectural features such as hybrid attention, LatentMoE and Engram, bringing the KV cache size down to just 1/24 that of Hy3 and 1/6 that of Hy4; in the post-training stage, it obtained comprehensive intelligence by self-distilling Hy4, then used reinforcement learning to cut end-to-end task output length by 50%; and in the inference stage, it combined megakernel, 4-bit quantization, DFly and other methods to push decoding speed beyond 150 tps.

The approach behind Hy4 flash is "extreme streamlining of activated parameters" — activating only 12B out of 224B total parameters, and directly compressing output length through reinforcement learning, so that the number of tokens consumed per task is itself smaller.

Tencent also invited 85 internal WorkBuddy users to conduct blind tests on 259 daily work tasks across 27 categories. The results showed that Hy4 flash's user satisfaction was close to that of Claude Haiku 5.5, its per-task cost was lower than all participating models, and its task duration was also the shortest.

In terms of overall effectiveness, Tencent said Hy4 flash is slightly better than Hy4 preview.

Two Products, Each Targeting Different Users

In terms of positioning, the two models do not entirely target the same type of user.

Qwen4-Max is more like a flagship product aimed at complex enterprise-level scenarios. In its launch materials, Alibaba listed several deployment cases: in an automotive manufacturing scenario, the manual input required for programming a single workpiece's processing was reduced from about 475 minutes to 24 minutes, a reduction of roughly 95%; in corporate IPO legal due diligence, the model processed 198 scattered working papers and delivered a complete 86-page report of more than 60,000 words; and in a financial investment research scenario, it can call 90 MCP tools to obtain market data on A-shares and Hong Kong and US stocks and generate traceable reports.

Hy4 flash, by contrast, focuses more on daily office efficiency scenarios. Tencent's recommended use cases include deep search, document processing, PPT generation, simple front-end displays and Computer Use, adding that "for complex tasks such as long-horizon software engineering, use Hy4, which has a higher intelligence ceiling."

In terms of accessibility, Hy4 flash is already live on products including Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis. Before November 30, users can use it free for a limited time on WorkBuddy, and the Tencent Cloud TokenHub API is at a limited-time 50% discount. Tencent also said it plans to open-source Hy4 flash and iterate on it frequently.

A Shift in Evaluation Methods

In this launch, both companies coincidentally emphasized internal evaluation sets that are "close to real work scenarios," rather than relying solely on public leaderboards.

Tencent built the WorkIndex evaluation set based on typical WorkBuddy tasks, "focusing on the output quality of daily work, single-round output length and end-to-end time, rather than public leaderboard scores far removed from daily work."

Alibaba's Work360 framework, meanwhile, covers 26 first-level domains, 182 second-level subdomains and 721 third-level细分 domains, attempting to systematically cover industrial and professional scenarios.

Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.

Comments

We need your insight to fill this gap
Leave a comment