AI News

Alibaba releases Qwen4-Max, Tencent Hunyuan releases Hy4 flash, with per-task costs down 58% and compressed to 1/8 of competitors, respectively

WallstreetcnUpdated Nov 16 21:06

AI summary

On November 16, Alibaba released Qwen4-Max, the first flagship model in the Qwen4 series, and Tencent Hunyuan launched the lightweight model Hy4 flash. According to Alibaba's official Work360 evaluation data, Qwen4-Max's average per-task cost dropped from 26.58 yuan to 11.16 yuan, a decrease of 58%; it has 2.4T total parameters, activates 71B per token, and supports native vision and a million-token context window. Tencent said Hy4 flash has 224B total parameters and only 12B activated parameters; in WorkIndex evaluations, its per-task cost is 1/20 that of Hy4 and 1/8 that of DeepSeek-V4.1-Flash-1017, and it has already launched on products including Tencent Cloud TokenHub, Yuanbao, and WeCom.

Why read this

AI score 79

The two models released on the same day each disclose parameter scale, token efficiency, and per-task cost, allowing a side-by-side look at the flagship and lightweight routes.

Topics

Tags

Body · AI translation

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

According to Alibaba's officially published Work360 evaluation data, the average cost per task for Qwen4-Max dropped from 26.58 yuan to 11.16 yuan, a decrease of 58%. Tencent, for its part, said that the per-task cost of Hy4 flash in the WorkIndex evaluation is only 1/20 that of Hy4, and only 1/8 that of DeepSeek-V4.1-Flash-1017 (max tier, busy-hour price).

It is worth noting that this week Alibaba's and Tencent's Q3 earnings reports are set to be released on November 18 and November 19, respectively.

Qwen4-Max: 2.4 trillion parameters, task cost cut in half

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

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

In terms of performance, Alibaba officially stated that compared with the previous generation Qwen3.8-Max, Qwen4-Max's score in the comprehensive Work360 evaluation rose from 47.74 points to 52.69 points, and it also showed significant improvements on multiple Agent-specific benchmarks such as DeepSWE, AA-Briefcase, and HLAB-Vals.

At the architecture level, Qwen4-Max upgrades the 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, enhancing cross-layer information transfer; and N-gram Embedding expands model capacity with relatively low additional computational overhead.

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

Hy4 flash: only 12B activated parameters, cost pressed down to 1/8 of competitors and 1/20 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." In terms of API pricing, input is 0.5 yuan per million tokens and output is 2 yuan, and only 0.02 yuan when the cache is hit. In the WorkIndex evaluation, the per-task cost is only 1/13 that of Hy4 preview, 1/20 that of Hy4, and 1/8 that of DeepSeek-V4.1-Flash-1017 (max tier, busy-hour price).

To achieve this cost level, Tencent carried out targeted optimizations in multiple stages. In the pre-training stage, architectural features such as hybrid attention, LatentMoE, and Engram were introduced, and the KV cache size is only 1/24 that of Hy3 and 1/6 that of Hy4; in the post-training stage, comprehensive intelligence was obtained through self-distillation from Hy4, and then reinforcement learning reduced the end-to-end task output length by 50%; in the inference stage, methods such as megakernel, 4-bit quantization, and DFly were combined to make the decoding speed exceed 150 tps.

Hy4 flash's approach is "extremely streamlined activated parameters"—only 12B of the 224B total parameters are activated, and reinforcement learning directly compresses the output length, so that the number of tokens consumed per task is itself smaller.

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

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

Two products, each targeting different users

From their positioning, the two models do not entirely target the same type of users.

Qwen4-Max is more like a flagship product aimed at enterprise-level complex scenarios. In its release materials, Alibaba listed multiple implementation cases: in an automotive manufacturing scenario, the manual input required for programming the machining of a single workpiece dropped from about 475 minutes to 24 minutes, a reduction of about 95%; in corporate IPO legal due diligence, the model processed 198 scattered working papers and delivered a complete report of 86 pages and more than 60,000 words; in a financial investment research scenario, it can call 90 MCP tools to obtain market data such as 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 display, and Computer Use, and "for complex tasks such as long-horizon software engineering, use Hy4, which has a higher intelligence ceiling."

In terms of accessibility, Hy4 flash has already launched on products including Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis. Before November 30, users can use it for 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 frequently.

Changes in evaluation methods

In this release, 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, single-round output length, and end-to-end time of daily work, rather than public leaderboard scores far removed from daily work."

English text is being prepared.

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the special investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Invest accordingly at your own risk.

Source

Wallstreetcn