What Hy4 Preview Is, in Plain Numbers

On August 28, 2026, Tencent released Hy4 Preview as open weights — 770 billion total parameters in a Mixture-of-Experts architecture that activates just 49 billion per token on each forward pass. The context window exceeds 1 million tokens. The license is Apache 2.0, meaning commercial use is unrestricted. The model is available on Hugging Face, Tencent’s own TokenHub platform, and OpenRouter.

The architecture features 78 backbone layers with 256 routed experts and borrows what Tencent calls “Gated DeepSeek Sparse Attention” combined with IndexCache — a mechanism for cross-layer KV-cache reuse that makes the 1M-context window economically viable to serve rather than just theoretically possible. This is the part of the spec worth paying attention to: context length without efficient serving is marketing.

For scale: the predecessor Hy3 had 295 billion total parameters, 21 billion active, and a 256,000-token context window. Hy4 Preview is not a minor iteration. The jump from 295B to 770B total parameters and from 256K to 1M+ context is the kind of generational gap that warrants treating this as a new model family rather than an update.

The Benchmark Claims (and Why to Read Them Carefully)

Tencent ran an internal evaluation across 203 engineering tasks judged by 163 company-selected experts scoring on a 0-to-4 scale. Hy4 Preview averaged 2.99; Kimi K3 scored 2.94; GLM-5.3 scored 2.92. That is a 0.07-point spread across the top three — performance parity, not a runaway lead. Tencent presented this as evidence of superiority. A more accurate reading is that these three models are clustered at roughly the same capability level.

On external agentic benchmarks, Hy4 scores 85.4 on Terminal-Bench 2.1 and 83.7 on MCP-Atlas, which measures multi-agent coordination over structured tool use. These are credible benchmarks, but no independent third party has yet reproduced Tencent’s full internal evaluation. The numbers are directional until that changes.

This pattern is not unique to Tencent. Kimi K3 — released by Moonshot AI on July 27, 2026, with 2.8 trillion total parameters under an MIT license — made similar claims before independent reviewers confirmed its 57.1 score on the Artificial Analysis Intelligence Index v4.1. You can read the full Kimi K3 benchmark analysis here. Hy4 deserves the same scrutiny before being slotted into production workflows.

A 1 Million-Token Context Window — With Caveats

The 1M-token context is Hy4’s most prominent technical feature. At that length, a single prompt can hold roughly 750,000 words — enough for multiple large codebases, extensive document collections, or long agentic conversation histories. For real-world use cases like cross-document analysis and codebase-level reasoning, this is meaningful capacity, not a benchmark curiosity.

But Tencent acknowledged two significant constraints. First, throughput: Hy4 generates approximately 36 tokens per second. For a chatbot, that is slow but usable. For a production agentic pipeline processing many concurrent requests, it is a serious bottleneck. Second, the model “over-verifies its own work” — Tencent’s own language — meaning it can spend excessive compute re-checking outputs rather than progressing through a task. In agentic settings, this compounds the throughput problem.

IndexCache does mitigate repeated-context serving costs meaningfully. Cached input tokens cost $0.042 per million — a 20× reduction from the base input price of $0.834 per million. For workflows that repeatedly reuse large system prompts or document collections, that pricing structure changes the cost model significantly. DeepSeek V4 Pro remains cheaper overall on uncached input, but the 20× cache discount could tip the math for specific agentic use cases.

Where Hy4 Sits in the Open-Weight Race

August 2026 was an unusually dense period for open-weight releases. Alibaba dropped Qwen3.8-Flash-Next and Z.AI launched GLM-5.3-Flash on August 26 — two days before Tencent’s Hy4 announcement. All three Chinese labs shipped within a 72-hour window. Whether coordinated or not, the timing reflects a broader pattern: major Chinese AI labs are now competing openly on published benchmarks and permissive licensing rather than keeping weights closed.

The strategic logic is straightforward. Open-source releases allow Chinese labs to drive global adoption and downstream fine-tuning rather than relying solely on proprietary API infrastructure. Tencent can afford to open-source Hy4 because the real monetization happens in its cloud products — WorkBuddy, CodeBuddy, Yuanbao, and ima — where Hy4 is already integrated and available free for two weeks from launch.

Compared to Western open-weight efforts, Hy4 sits above Mistral Large 3 in raw parameter count and context window, though a fair performance comparison requires independent evaluation. Meta’s Llama 4 family targets smaller-footprint efficiency at a different deployment tier — Hy4’s minimum viable deployment requires H200-class hardware or a multi-node cluster, which excludes most teams without cloud infrastructure.

What This Means for Developers

The practical barrier is the 770-gigabyte checkpoint. Self-hosting Hy4 Preview requires eight or more H200-class GPUs or a multi-node cluster. FP8 quantized weights are available alongside full-precision versions, with deployment support for vLLM and SGLang. But even quantized, this is not a model you spin up on consumer hardware. For most developers, Hy4 is an API-first experience.

Via OpenRouter, the $0.834 per million input tokens is competitive for the capability tier. The 20× cached-input discount is the figure worth modeling if you are building RAG systems or multi-turn agentic pipelines that repeatedly reuse long context. Two weeks of free access on WorkBuddy and CodeBuddy gives teams a no-commitment window to test the model on real tasks before committing to API costs — which is the right evaluation sequence before architecting a pipeline around 1M-context claims.

Hy4 Preview is a preview label for a reason. Tencent explicitly listed over-verification and low throughput as known issues, signaling that a stable release will address them. A Hy4 stable with 60+ tokens per second and improved agentic reliability would be a meaningfully different product. Until then, the benchmarks are promising but the deployment constraints are real.

Further Reading

• Tencent open-sources 770B Hy4 preview model — TechNode’s original coverage with release timing and product integration details
• Hy4 Preview: 770B Open Weights at $0.83/M Input — technical breakdown of the IndexCache architecture, pricing tiers, and benchmark methodology
• Tencent’s Hunyuan Hy4 and China’s Open-Model Race — broader context on why Chinese labs are converging on open-weight releases as a competitive strategy