Tencent Launches Hy4 Flash for Low-Cost AI Inference

Tencent Hy4 Flash: Affordable Enterprise AI Model Tencent Hy4 Flash: Affordable Enterprise AI Model

Tencent is targeting the economics of enterprise artificial intelligence with Hy4 Flash, a new large language model designed to deliver advanced reasoning and productivity capabilities at a fraction of the cost of larger models. The latest addition to the Tencent Hy model family combines a 224-billion-parameter architecture with just 12 billion active parameters, multimodal input support and a context window of up to one million tokens. The launch highlights an increasingly important competitive priority in generative AI: making powerful models affordable and fast enough for everyday business applications.

Tencent Hy4 Flash Targets the Economics of Enterprise AI

The race to build more capable artificial intelligence models is increasingly becoming a race to make them cheaper to operate. As enterprises move from conversational assistants to AI agents that perform research, edit documents and execute multistep workflows, inference costs and response times are becoming important factors in model selection.

Tencent’s Hy4 Flash reflects this shift. Rather than competing exclusively on model size or benchmark rankings, the company is emphasizing practical productivity, lower token consumption and faster inference.

The new model has 224 billion total parameters, including 64 billion Engram parameters, but activates only 12 billion parameters during processing. It supports multimodal inputs and a context window of up to one million tokens, positioning it for applications that require processing large documents, visual information and extended work sessions.

A smaller active footprint with a larger architecture

Hy4 Flash uses a combination of hybrid attention, LatentMoE (Latent Mixture of Experts) and Engram architectural features. These techniques are intended to reduce computational requirements while preserving the capabilities of a substantially larger model.

One important optimization concerns the key-value (KV) cache, which stores information from previous tokens to support efficient inference. Tencent says Hy4 Flash’s KV cache is approximately 24 times smaller than Hy3’s and six times smaller than Hy4’s.

Reducing KV cache requirements can help lower memory consumption, improve hardware utilization and make long-context inference more economical.

The model also inherits elements of Hy4’s architecture through its post-training process. Tencent says reinforcement learning optimization reduces average end-to-end task output tokens by 50%, another mechanism for controlling inference expenditure.

The result is a model designed around a different efficiency equation: retaining a relatively large total parameter architecture while limiting the number of parameters activated for each processing step.

Pricing puts pressure on the AI inference market

Hy4 Flash’s pricing is one of its most prominent selling points.

Hy4 Flash API pricing

$0.070

Input / 1M tokens

$0.281

Output / 1M tokens

$0.003

Cached input / 1M tokens

Standard announced prices in US dollars, before promotional discounts. Source: Tencent.

These rates make Hy4 Flash an inexpensive option for applications involving frequent model calls, repeated context and large volumes of generated content.

Tencent’s internal WorkIndex benchmark provides another perspective on the economics. Covering PowerPoint creation and editing, document processing, spreadsheet repair, audio/video processing and visual understanding, the benchmark measures task completion costs alongside output quality and turnaround time.

Tencent reports that Hy4 Flash delivers a per-task cost approximately 13 times lower than Hy4 Preview, 20 times lower than Hy4 and eight times lower than DeepSeek-V4.1-Flash-1017 at its Max tier during peak-hour pricing.

The company also says Hy4 Flash slightly outperforms the larger Hy4 Preview in overall effectiveness. These are company-reported internal benchmark results, not independently reproduced comparisons, and should be evaluated accordingly.

Measuring AI through completed work

One notable aspect of the launch is Tencent’s emphasis on practical task evaluation rather than relying exclusively on public model benchmarks.

The company developed WorkIndex, an internal evaluation framework based on tasks commonly performed through its WorkBuddy productivity platform. It considers deliverable quality, single-turn output length and task completion time.

Tencent also conducted a blind evaluation involving 85 internal WorkBuddy users, covering 259 tasks across 27 categories.

According to the company, Hy4 Flash achieved user satisfaction close to Claude Haiku 5.5 while delivering the lowest per-task cost and shortest completion time among the models evaluated.

This approach reflects a broader change in enterprise AI evaluation. A model’s usefulness increasingly depends on whether it can produce a usable spreadsheet, prepare a presentation, repair a document or complete a computer-based workflow within acceptable time and cost limits.

However, the evaluation involved Tencent employees and internally selected tasks. Independent testing across different industries, workloads and user groups will be important for establishing how well the model performs outside Tencent’s ecosystem.

Inference optimization drives speed and affordability

Beyond model architecture, Tencent has optimized the inference pipeline using megakernel techniques, 4-bit quantization and DFly.

The company reports decoding speeds exceeding 150 tokens per second.

These optimizations address a central challenge for large language models: computational efficiency does not depend exclusively on the number of active parameters. Memory bandwidth, hardware utilization, quantization and inference software also influence the cost and speed of generating responses.

For enterprises building AI automation platforms and autonomous agents, these improvements could matter considerably. Agentic applications often involve repeated model calls, intermediate reasoning, tool interactions and multiple stages of execution. Even relatively inexpensive individual requests can become costly when multiplied across complex workflows.

Gartner projected in August 2026 that AI inference costs per agentic workflow could increase more than fivefold through 2028, despite improvements in underlying model economics. Gartner This illustrates why lower token prices alone may not solve the overall cost problem.

Hy4 Flash’s combination of reduced active parameters, smaller KV cache requirements and faster decoding directly addresses several parts of that equation.

Tencent expands its AI model ecosystem

Hy4 Flash is available through Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter integration planned.

Tencent is also offering free access through WorkBuddy until November 30, 2026, alongside a 50% discount on standard pricing through Tencent Cloud TokenHub during the promotional period.

The availability across multiple products gives Tencent an opportunity to connect its foundation models with productivity applications, enterprise communication tools and consumer-facing AI services.

The strategy also places the company in competition with international model providers such as Google, Microsoft and OpenAI, as well as Chinese AI developers including DeepSeek.

Tencent recommends Hy4 Flash for everyday productivity workloads, including deep research, document processing, presentation generation, simple front-end development and computer use. More demanding, long-horizon software engineering tasks remain better suited to the larger Hy4 model.

Open-source ambitions and the next phase of competition

Tencent says it plans to open-source Hy4 Flash and introduce frequent updates to improve its capabilities and efficiency. However, the company has not established a public release date for the model weights in its announcement.

That distinction matters for developers evaluating the model. API availability provides immediate access, but open weights could offer additional flexibility for organizations seeking self-hosted deployment, infrastructure control and customization.

The broader significance of Hy4 Flash is its emphasis on practical AI economics. As enterprises deploy more agents and automated workflows, the competitive advantage may increasingly belong to models that can deliver reliable results at predictable operating costs.

Tencent is betting that a balance of intelligence, speed and affordability will make Hy4 Flash a practical choice for everyday AI workloads. Its longer-term position will depend on independent performance testing, open-source execution and adoption beyond the company’s own product ecosystem.

Market Landscape

The launch reflects five major developments across the global AI industry.

1. Inference economics is becoming a competitive differentiator

Model providers are competing on cost per completed task, not simply cost per million tokens. Lower inference costs make high-volume enterprise automation more commercially viable.

2. Mixture-of-Experts architectures are gaining importance

Architectures that activate only selected parameters can reduce computational requirements while retaining substantial model capacity. Tencent’s Hy4 Flash is an example of this approach.

3. AI agents are changing model requirements

Research, spreadsheet editing, document creation and computer-use workflows require more than conversational fluency. They demand reliable tool execution, multimodal understanding and efficient long-context processing.

4. Open-source models are challenging proprietary AI platforms

Tencent’s planned open-source release could strengthen competition with proprietary offerings from OpenAI, Google and Anthropic, although deployment flexibility will depend on the eventual licensing and weight availability.

5. Enterprise AI infrastructure is becoming more important

Inference optimization, quantization, memory efficiency and model routing are becoming central to enterprise AI infrastructure strategies.

Gartner’s forecast of a more than fivefold increase in inference costs per agentic workflow through 2028 highlights the importance of these developments. Gartner

Top Insights

  • Tencent Hy4 Flash combines 224 billion total parameters with only 12 billion active parameters, targeting lower-cost enterprise AI inference.
  • The model supports multimodal inputs, a one-million-token context window and reported decoding speeds exceeding 150 tokens per second.
  • Tencent claims up to 20 times lower task costs than Hy4 based on its internal WorkIndex benchmark.
  • Blind testing involving 85 internal users found satisfaction comparable to Claude Haiku 5.5, according to Tencent.
  • Planned open-source availability could expand developer adoption, although Tencent has not announced a specific weights release date.

Power Tomorrow’s Intelligence — Build It with TechEdgeAI

Grow Your
Brand Visibility

Looking to publish a press release, guest article, interview or podcast? Connect with us.

GET FEATURED
Subscribe

Sign up today for exclusive insights and updates.

Newsletter Signup