Tencent Open-Sources 770B-Parameter Hy4
๐กA new open-source 770B model brings a million-token context window to developers.
โก 30-Second TL;DR
What Changed
Has 770 billion total parameters and 49 billion activated parameters
Why It Matters
Hy4 adds another very large open-source model option for developers building long-context applications. Its mixture of high total capacity and lower activated parameters could be relevant for teams evaluating quality, context length, and serving efficiency together.
What To Do Next
Test Hy4 through its API on a long-document or code-repository workload and compare output quality, latency, and serving cost with your current model.
Key Points
- โขHas 770 billion total parameters and 49 billion activated parameters
- โขSupports a context window exceeding 1 million tokens
- โขAvailable through WorkBuddy, CodeBuddy, Yuanbao, and ima, with API access
๐ง Deep Insight
Background and context from public sources โ not the original article. 9 sources cited.
๐ Enhanced Key Takeaways
- โขThe model is released under the Apache 2.0 open-source license, facilitating broader adoption in commercial and research environments.
- โขHy4 utilizes a 10B parameter Multi-Token Prediction (MTP) layer specifically designed to enhance speculative decoding performance.
- โขInternal blind evaluations against industry peers showed Hy4 achieving a 2.99/4.00 score, surpassing GLM-5.3 and Kimi K3 in engineering-focused tasks.
- โขTencent reported a 31.8% increase in end-to-end training and inference throughput by integrating Hy4 into their internal development workflows.
- โขThe model is available for international developers via Tencent Cloud TokenHub and OpenRouter, with pricing set at $0.834 per million input tokens.
๐ Competitor Analysisโธ Show
| Feature | Hy4 (Tencent) | GLM-5.3 | Kimi K3 |
|---|---|---|---|
| Total Parameters | 770B | N/A | N/A |
| Activated Parameters | 49B | N/A | N/A |
| Context Window | >1M tokens | N/A | N/A |
| Expert Eval Score | 2.99 | 2.92 | 2.94 |
๐ ๏ธ Technical Deep Dive
- Architecture: Mixture-of-Experts (MoE) design with 770B total parameters and 49B activated parameters.
- Speculative Decoding: Incorporates a 10B parameter Multi-Token Prediction (MTP) layer to accelerate inference.
- Tooling: Includes specialized parsers optimized for complex tool use and multi-step reasoning.
- Training Data: Curated datasets focused on software engineering, game development, finance, and security.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
โณ Timeline
๐ Sources (9)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: TechNode โ
This is a summary, not the original. Read the source, or get the weekly briefing.
Weekly AI briefing
One email a week. Unsubscribe anytime.