Hy4: A New Open-Source Flagship with 770B Parameters, 1M Context, and Multimodal Capabilities

Background: Large Model Competition Enters a Deeper Phase

As large language model technology evolves rapidly, the competition between open-source and closed-source approaches has intensified. Vendors are no longer simply chasing parameter scale; they are placing greater emphasis on real-world performance and developer experience. The recently released Hy4 has drawn significant industry attention due to its unprecedented parameter volume and expanded capabilities. At the same time, third-party benchmark results show subtle discrepancies with official claims. This gap reflects the complexity of current LLM evaluation systems and the diversity of model behavior across different testing dimensions.

Model Architecture and Core Features

1. Scaling Up: From 300B to 770B

Hy4 uses a Mixture of Experts (MoE) architecture with 770B total parameters, but only 49B parameters are activated per inference. Compared with its predecessor, which had under 300B parameters, the model size has grown by roughly 2.6x, making it the largest model in the series to date. This “large but sparse” design philosophy increases model capacity while keeping inference costs under control — a classic trade-off in modern LLM engineering.

The core value of the MoE architecture lies in specialization: different expert networks handle different types of input patterns. This explains why inference efficiency remains practical despite the massive parameter growth. For developers, understanding this mechanism helps in estimating model performance and resource consumption for specific tasks.

2. On-Device Capability Breakthrough: Million-Token Context Window

Hy4 supports a context length of 1 million tokens, moving well beyond the “long context” category into true “ultra-long context” territory. This capability delivers value in two typical scenarios: analyzing large code repositories where the model can understand an entire project’s structure at once rather than reading scattered file fragments, and processing long documents such as complete technical manuals or research reports.

Long-context capability is not simply a matter of stretching the window size. It involves systematic engineering work including attention mechanism optimization, positional encoding improvements, and adjustments to training data distribution. In practice, developers should pay attention to the model’s information recall rate in long contexts — whether it can accurately reference key details that appeared early in the context, rather than merely “seeing” them.

3. Multimodal and Embodied Intelligence: From Reading to Seeing to Acting

The official release introduces two major capability upgrades over the preview version:

• Visual understanding: The model supports image input for the first time, enabling it to parse charts, screenshots, hand-drawn sketches, and other visual information. It can now “read” architecture diagrams, UI prototypes, and data visualizations, then provide analysis or generate code based on visual content.
• Computer use: The model can execute cross-application tasks by combining command-line and graphical interface operations. This connects the pipeline of language understanding, task planning, and tool invocation, giving the model a form of embodied intelligence that simulates basic human computer operations.

A fitting analogy: the preview model was like a scholar who could only read text, while the official release has grown “eyes” (vision) and “arms” (operation), moving from single-modal text understanding toward multimodal interaction and task execution.

Pricing and Access

In terms of commercial positioning, Hy4 continues a budget-friendly pricing strategy. API pricing remains unchanged from the previous generation: 6 CNY per million input tokens, 18 CNY per million output tokens, and 0.3 CNY per million cached tokens. For comparison, current flagship-tier models such as Kimi K3.1 Pro charge 10 USD per million output tokens (approximately 70+ CNY), making Hy4 roughly a quarter of that price. This pricing clearly targets developers who want flagship-level performance at a lower cost.

Access is straightforward: select “Hy4” from the model dropdown in WorkBuddy or CodeBuddy. The current version is in a limited free period, with daily free credits available for users to experience the full capabilities. The “thinking intensity” option can be adjusted between high, medium, and low settings to balance response speed and output quality.

Interpreting Benchmark Data: Official Scores vs. Third-Party Results

1. Understanding the Gap in Comprehensive Intelligence

The official claim of “the strongest open-source model” coexists with third-party composite scores showing an 11-point gap behind a competitor such as Kimi K3.1 Pro. This phenomenon deserves closer analysis. The discrepancy may stem from several factors:

• Different benchmark compositions: Evaluation systems weight the definition of “intelligence” differently. Some emphasize mathematical reasoning, others lean toward code generation, and still others focus on Chinese language understanding. A composite intelligence index is a weighted aggregate, and performance in specific domains may vary significantly.
• Hallucination rate measurement: The “high hallucination rate” noted by third-party evaluators may reflect the model’s performance on specific factual question-answering tasks, which does not necessarily conflict with reasonable generation in creative writing scenarios. Hallucination control remains a common challenge for all LLMs, and evaluations should be interpreted in the context of specific use cases.
• Parameter count vs. activated parameters: In MoE architectures, the difference between total and active parameters is substantial. Inference configurations across different evaluation environments — such as quantization precision and parallelism strategies — can also affect results.

2. A Practical Approach to Benchmark Numbers

Benchmark numbers are signposts, not destinations. When selecting a model, developers should focus on how closely their business scenarios align with the benchmark tasks. For example, if the workload involves heavy structured data analysis, benchmarks weighted toward mathematical reasoning are more relevant. If the use case is open-domain conversation, language naturalness and contextual consistency matter more. A practical approach is to treat official scores as an upper-bound indication of capability, treat third-party evaluations as general references across different dimensions, and ultimately rely on custom test sets tailored to your actual workloads.

Hands-On Case Studies: From Theory to Practice

Case 1: 3D Scene Generation — The “Gold Miner” Challenge

• Test objective: Evaluate the model’s code generation, complex logic orchestration, and multi-component collaboration capabilities.
• Test method: Provide Hy4 with a natural language description asking it to generate a playable “Gold Miner” style 3D mini-game using common stacks such as JavaScript and Three.js. The competitor model Kimi K3.1 Pro was given the same task.
• Observed performance:
  • Hy4 handled the complete loop of “hook launch — mineral grabbing — retrieval” well, generating clear code architecture with reasonable module boundaries.
  • In edge case handling — such as hook retraction logic when the hook exceeds screen boundaries and the dynamic relationship between mineral weight and pull-back speed — Hy4 performed more precisely than the competitor. It also unexpectedly added extra interactive elements like “bomb rocks,” demonstrating good extensibility.
  • The generated code ran directly with only minor styling issues, while the core gameplay mechanics were fully functional.

Case 2: Hand-Drawn Sketch to Web Page — Combining Vision and Syntax

• Test objective: Assess the model’s visual understanding and the accuracy of cross-modal conversion from visual information to natural language description to code implementation.
• Test method: Provide a hand-drawn wireframe of a web page containing a navigation bar, main content area, sidebar, and footer. Ask Hy4 to generate the corresponding HTML/CSS code based on the image.
• Observed performance:
  • Hy4 accurately identified the positional relationships of all core regions and mapped them to CSS Flexbox or Grid layout structures.
  • Visual weight judgment for text hierarchy (headings vs. body text) was accurate, with generated font sizes and color contrast largely matching the design intent.
  • One interesting detail: Hy4 noticed the “irregular” lines in the hand-drawn sketch and inferred that the area was likely an image carousel component, reserving an appropriate container in the code. This ability to reason from imperfect visual input is a strong indicator of robust visual understanding.

Target Users and Practical Recommendations

Considering the combination of price, capability, and openness, Hy4 is best suited for three groups:

Target User	Typical Needs	Recommendations
Independent developers	Rapid prototyping, utility code generation	Take advantage of the free trial period to test code generation. Use the “medium” thinking intensity setting to balance speed and quality
Small to mid-sized engineering teams	Lower API costs, potential self-hosted deployment	Focus on the million-token context capability for code repository semantic analysis or document Q&A systems
AI application researchers	Exploring multimodal interaction and agent task orchestration	Combine visual understanding with computer-use capabilities to build a complete “read — plan — execute” agent pipeline

It is worth noting that Hy4 still has gaps compared with top-tier closed-source models in long-form text generation and factual accuracy in knowledge-based Q&A. If your use case demands high factual reliability — such as medical or legal consultation — consider adding RAG (Retrieval-Augmented Generation) or a knowledge base verification layer.

Summary and Outlook

The core value of the official Hy4 release lies in offering a flagship-level configuration — a large-parameter MoE architecture, million-token context, multimodal perception, and computer-use capabilities — at an extremely competitive price point. Its release marks a step forward for open-source models in overall capability ceilings and signals that cost-performance competition has entered a new phase.

For developers, there is no need to be swayed by “open-source number one” claims, nor discouraged by lower scores on particular benchmarks. The pragmatic approach is to design multi-dimensional validation sets based on your own scenarios and test the model on the capabilities most relevant to your business. In engineering, what works best is what matters most. Hy4 provides a sufficiently broad space for exploration, and it is worth investing time to understand what it can do for your specific use cases.