On November 16, 2026, the Tencent Hunyuan team officially released Hy4 Flash, positioning it as “the most cost-effective Hy model to date.” The model has 224B total parameters, including 64B Engram parameters, with only 12B parameters activated per token, and supports native multimodality and a 1M-token context length. According to the price sheet published by Tencent Cloud TokenHub, API input is priced at $0.070 per million tokens and output at $0.281 per million tokens—compared with the full Hy4 ($0.834 input, $2.501 output), input is down about 92% and output about 89%.
Tencent cites internal evaluation data claiming that Hy4 Flash’s user satisfaction is close to Claude Haiku 5.5 while cost per task is about 18 times lower. All performance comparison data comes from tests organized internally by Tencent; as of the release date, no third-party laboratory has independently reproduced them.
Three-Stage Collaborative Optimization: Where Efficiency Comes From
Hy4 Flash’s low cost is not achieved simply by compressing model capability in exchange for savings; it is the result of stacked technical decisions across pretraining, post-training, and inference.
In pretraining, the Hunyuan team introduced a combination of hybrid attention, LatentMoE (a sparse expert framework), and Engram (a scalable lookup-table-based parameter module), reducing KV cache size to 1/24 that of Hy3 and 1/6 that of Hy4. The direct role of the KV cache is to store previously computed attention key-value pairs for reuse—a smaller cache means less memory usage when handling long contexts and lower large-scale deployment costs, which is especially critical for million-token context scenarios.
In post-training, the team used self-distillation from the full Hy4 so the model inherits a strong level of general intelligence, then used reinforcement learning to compress average end-to-end task output length by 50%. Shorter outputs directly translate into lower output-token costs in batch-task scenarios, and the unit price of output tokens is usually several times that of input tokens.
In inference, Tencent reports that megakernel fusion, 4-bit quantization, and DFly scheduling optimization push decoding speed above 150 tokens/s, supporting large-scale low-cost deployment and operations. This speed figure is an official self-measured value; according to third-party platform data, Claude Haiku 5.5’s median generation speed on OpenRouter is about 171 tokens/s, higher than the currently observable Hy4 Flash value, but Hy4 Flash’s OpenRouter access had not gone live at release, so this comparison has limitations.
Complete Data from the Internal Blind Test: What Tencent Selectively Presented
Tencent organized 85 internal WorkBuddy users for a blind test covering 259 daily work tasks across 27 categories. The test covered 8 evaluation sets, including slide generation and editing, document processing, spreadsheet repair, and audio and video understanding, and built an evaluation system called WorkIndex that focuses on output quality, single-turn output length, and task completion time rather than public leaderboard scores.
Tencent’s main external conclusion is: Hy4 Flash has a user satisfaction score of 2.79, close to Claude Haiku 5.5’s 2.81; cost per task is only ¥0.26, far below Claude Haiku 5.5’s ¥4.73, a gap of about 18 times; and task completion time (median 6.7 minutes) is also the shortest among the compared models.
However, in the complete results of the same test, DeepSeek-V4.1-Flash-1017’s user satisfaction score reached 3.09, higher than both Hy4 Flash and Claude Haiku 5.5, and this figure was not included in Tencent’s core external messaging. This does not mean Hy4 Flash’s test results are fabricated, but it suggests that the statement “performance close to Haiku” presupposes specific task scenarios and a specific evaluation framework, and under the same framework, DeepSeek has already produced a better cost-performance combination.
The gap on complex coding tasks is more obvious. According to Terminal-Bench 4.0 test data, Hy4 Flash scored 19.2, while full Hy4 scored 44.4 and DeepSeek V4.1 Flash 1017 scored 40.0. Tencent also explicitly recommends that complex tasks such as long-horizon software engineering still be handled by full Hy4, and that Hy4 Flash is designed for high-frequency, medium-complexity daily productivity work.
Industry Impact: Developers, Enterprises, and the Competitive Landscape
For independent developers, Hy4 Flash’s pricing directly changes the cost equation for high-frequency call scenarios. In batch tasks that do not require deep reasoning, such as document processing, meeting minutes, and spreadsheet repair, the $0.07/M input price makes previously cost-sensitive product forms viable. Tencent also offers a cached input option at $0.003/M (compared with Hy4’s $0.042/M); for application-layer developers that frequently use fixed system prompts, the actual blended cost will be further compressed.
For enterprise users, model selection logic needs layered consideration. Hy4 Flash is distributed through Tencent’s existing product lines (WorkBuddy, WeCom, QQ, ima, Yuanbao), and its deployment path is deeply tied to the Tencent ecosystem. For enterprises already building workflows on Tencent Cloud, Hy4 Flash offers a low-friction cost-reduction path; for cross-cloud or multi-vendor enterprises, truly comparable horizontal testing will only be possible after OpenRouter access.
From a competitive landscape perspective, Hy4 Flash’s release further transmits price pressure from the domestic large-model lightweight inference tier to the international market. DeepSeek previously established a price reference in the global developer community with low prices and high performance, and Hy4 Flash’s entry increases the number of competitors in this tier. But DeepSeek V4.1 Flash 1017 achieved the highest satisfaction score in Tencent’s own test, showing that in overall cost-performance rankings, price is not the only variable.
Historical Comparison: The Credibility of the Lightweight Flagship Route
From a technical route perspective, Hy4 Flash follows the same line as the “Flash/Haiku/Mini” strategies of major labs in recent years: starting from flagship-model knowledge distillation, compressing per-token cost to below an order of magnitude of the flagship through architecture sparsification and inference optimization, in exchange for large-scale deployment feasibility. This route itself has been supported by the implementation of multiple products, and its technical logic is credible.
Such claims that a “lightweight version matches flagship performance” have historically often held in specific task domains, with significant discounts on reasoning-intensive or long-horizon tasks. Hy4 Flash’s Terminal-Bench data (19.2 vs 44.4) confirms this pattern, and Tencent did not avoid this gap; it clearly delineated the applicable scenario boundaries of the two products.
Outlook: Where Is the Validation Window
At present, all of Hy4 Flash’s performance data comes from inside Tencent. There are two key nodes for external validation: first, after OpenRouter access goes live, third-party developers will be able to compare it horizontally with other models under real workloads at a unified endpoint; second, Tencent plans to open-source the Hy4 Flash weights, and if open sourcing is delivered on time, academia and independent evaluators will obtain complete benchmarking conditions.
Before these two nodes arrive, the main signals are real user voting results on public leaderboards such as Lmarena or SEAL, independent evaluation scores on complex tasks (code generation, multi-hop reasoning), and the stability of 1M-token context in real application scenarios. The theoretically supported context length and the context length that remains usable under high load have historically not always been equivalent.
From a price-war perspective, this is already an irreversible structural change. Regardless of how third-party evaluations of Hy4 Flash turn out, the existence of $0.07/M input has itself recalibrated developers’ price expectations for this capability tier. For AI application-layer companies that rely on API-cost-based pricing models, this means cost models need to be recalculated; for end users, it means the same budget can support another order of magnitude of call volume.
Tencent says it plans to open-source Hy4 Flash and maintain high-frequency iteration. If this commitment is fulfilled, it will be the most direct basis for subsequently evaluating its real technical competitiveness. Before then, the claim “comparable to Claude Haiku 5.5” should be understood as: a test conclusion within Tencent’s custom office-task framework with internal users as evaluators, not a general capability equivalence claim certified by an independent institution.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接