Tencent has released Hy4 flash, a budget model in the Hy4 series for everyday work tasks: it is a 224B parameter MoE architecture where 12B parameters are activated per token, including 64B parameters for Engram conditional memory. The model is natively multimodal, accepts up to 1M tokens of context, and generates text at speeds exceeding 150 tokens per second, with API rates significantly lower than the flagship Hy4. According to Tencent's own measurements, the average task cost is 8–20 times lower than comparable models, while in a blind test involving 85 WorkBuddy users across 259 tasks, satisfaction was comparable to Claude Haiku 5.5. The model is already available: it can be tried for free in WorkBuddy until November 30, and a 50% discount applies to the API on Tencent Cloud TokenHub.


What happened
On November 16, 2026, Tencent introduced Hy4 flash—a new budget tier in the Hy4 lineup. The model is natively multimodal, supports up to 1M tokens of context, and generates text at over 150 tokens per second. API rates per 1M tokens: $0.003 for cached input, $0.070 for input, and $0.281 for output, whereas the flagship Hy4 costs $0.042, $0.834, and $2.501 for the same categories. The company demonstrates quality through its own blind test: 85 WorkBuddy users compared models across 259 tasks, and satisfaction with Hy4 flash was comparable to Claude Haiku 5.5. The average task cost in these measurements was 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 when converted to peak prices of its Max tier. The model lags in complex programming and long agent scenarios, and for such tasks, Tencent recommends the flagship Hy4 from the start.
Context
The cost-effectiveness of Hy4 flash is explained by a set of engineering solutions, not just pricing choices. The foundation is an aggressively sparse MoE architecture: with 224B parameters, only 12B are activated per token, meaning roughly five percent of the network. A separate 64B parameters are allocated for Engram—a conditional memory mechanism that extracts frequent patterns from a table instead of recomputing them. Other components of the stack also contribute to cost reduction: the LatentMoE module, hybrid attention with a KV-cache six times smaller in volume than the flagship Hy4, and 4-bit quantization during inference, while the claimed generation speed is provided by optimized megakernels along with the DFly technique. The release idea is to shift the focus from price per token to price per task: the publication in Tencent's official blog is titled “Introducing Hy4 flash: Ultimate cost-efficiency for everyday productivity,” and the model is positioned from the start as a cheap layer specifically for routine tasks where operation cost matters, not benchmark records.
Why this matters for the industry
For the industry, Hy4 flash raises the bar in the “good enough” intelligence segment. Mass office and agent scenarios with many similar requests, which previously did not work out in unit economics, become profitable at the stated rates, and a cheap API ceases to be a moat for wrappers around others' flagships. Price pressure in the segment of affordable agent models is now directed at DeepSeek, Alibaba, and Anthropic, who will have to respond to Tencent's multiple reduction in task cost. The next step depends on the promised opening of weights: if it happens, the community will get a cheap sparse 224B parameter MoE for self-hosting and distillations, and the lower threshold for entering agent infrastructure will drop even further. In such a setup, the “flash model for routine plus flagship for complex tasks” template may become the production standard, and competition may shift from price per token to memory quality and reliability of long agent chains—however, this is a forecast, not a completed fact.
Why this matters for users
The model is available to users right now: it works in Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis. Until November 30, it can be used for free in WorkBuddy, and a 50% discount applies to the API on TokenHub; its appearance on OpenRouter is promised but not yet confirmed, so access outside the Tencent ecosystem will require working with a Chinese cloud. Hy4 flash is worth trying where speed and price matter: office tasks, deep search, slide creation, and simple frontend demos. A practical check for a developer is to run a dozen or so real tasks of their product through the model and compare quality and actual cost with the current flash model in their own stack.
What is still unknown / limitations
The main source of figures is Tencent itself. The satisfaction blind test is a subjective vendor metric: the task distribution, prompts, evaluation criteria, and statistical significance are unknown, and Claude Haiku 5.5 metrics in the same setup have not been published. The headline claim that task costs are 18 times lower than Claude Haiku 5.5 cannot be verified with open data: this is a composite task cost metric from Tencent's product test, not a pure benchmark comparison. The task cost comparison with DeepSeek-V4.1-Flash-1017 is based on peak prices of the Max tier. The model weights are currently closed: opening is announced but not implemented, and the impact of 4-bit quantization on quality is not documented in detail.
Sources
- Introducing Hy4 flash: Ultimate cost-efficiency for everyday productivity — Tencent Hunyuan official blog
- Tencent Hunyuan released Hy4 flash — ChainCatcher
- Hy4 flash — Tencent Cloud TokenHub official product page
Author
Look at AI, editorial team
