Tencent Releases Hy4 Preview: 770B-Parameter MoE Model with 1M Context for Coding Agents

Tencent has released Hy4 preview, a mixture-of-experts (MoE) language model built for coding agents, tool-use workflows, and multi-step productivity tasks. The model is now listed on OpenRouter under the identifier tencent/hy4-preview.

Architecture and specs

Hy4 preview uses a mixture-of-experts design with 770 billion total parameters and 49 billion active parameters per forward pass — a ratio that keeps inference cost closer to a mid-sized dense model while retaining the capacity of a much larger network. The model supports a 1 million token context window, putting it in the same tier as long-context leaders like Gemini 1.5 Pro and Qwen2.5-1M.

According to Tencent, the model is specifically designed for:

• Coding agents
• Complex tool-use workflows
• Tasks requiring planning, context continuity, and sustained multi-step execution

No benchmark scores (MMLU, HumanEval, SWE-bench, or similar) have been published alongside this release. No training data cutoff date has been disclosed.

Pricing and availability

Hy4 preview is priced at $0.834 per 1M input tokens and $2.501 per 1M output tokens through OpenRouter. The listing shows a release date of August 28, 2026, and the model has been live for approximately three days at time of writing, with uptime and throughput data still insufficient for reporting. No dedicated app usage data is yet available on the platform.

As a "preview" release, Hy4 is explicitly a pre-final build. Tencent has not stated a timeline for a stable, non-preview version, nor has it disclosed whether pricing or the parameter configuration will change before general availability.

What this means

Hy4 preview extends Tencent's push into large-scale MoE architectures aimed squarely at agentic use cases rather than general chat. The 49B/770B active-to-total ratio signals a deliberate bet that sparse activation can deliver near-frontier capability at a fraction of the inference cost — a pattern also seen in DeepSeek's V3/R1 line and Alibaba's Qwen3-MoE releases, both of which compete directly in the Chinese large-model market.

The 1M token context window matters most for the use cases Tencent is targeting: coding agents that need to hold entire codebases in context, and tool-use chains that accumulate long histories of function calls and intermediate results. Without published benchmarks, it's impossible to verify how Hy4 preview stacks up against established coding models like Claude Sonnet, GPT-4.1, or DeepSeek-V3 on tasks like SWE-bench or LiveCodeBench — claims about its suitability for "sustained multi-step execution" remain Tencent's own framing until independent evaluations appear.

The pricing — roughly comparable to mid-tier frontier models — positions Hy4 preview as a cost-competitive option for developers building agentic pipelines, provided the model's real-world performance holds up. Given the preview label and the unusual forward-dated release listing, expect further revisions before Tencent commits to a stable release.