← Back to blog
Hy4 Preview: Specs, Benchmarks, Pricing, and Who Should Use It

Hy4 Preview: Specs, Benchmarks, Pricing, and Who Should Use It

Emma Reed

By Emma Reed

MyClaw Editorial

Hy4 Preview

Put Hy4 Preview to Work

Test your own prompts, files, and agent workflows with Hy4 Preview in MyClaw.

Chat with Hy4 Preview

AI Takeaway

  • What is Hy4 preview? Tencent's open-weight Mixture-of-Experts model has 770B total parameters, activates 49B per token, and supports a 1M-token context window.
  • What is it built for? Tencent targets long-horizon coding, tool-using agents, office analysis, game development, and scientific work.
  • How can you access it? Launch options include Tencent Cloud TokenHub, OpenRouter, or self-hosting the released model weights, including the FP8 variant.
  • What is the main caution? This is an early preview. Tencent warns that it can spend too long reasoning and over-verifying its work, while independent reliability evidence remains limited on release day.

What Is Hy4 Preview?

A 770B MoE Model With 49B Active Parameters

Hy4 preview is the Tencent Hy team's new flagship language model, released on August 28, 2026. It uses a Mixture-of-Experts architecture: the full model contains 770 billion parameters, but only 49 billion are activated for each token.

The 78-layer backbone uses 256 routed experts plus a shared expert in most layers, selects eight routed experts per token, and adds a native multi-token prediction layer for speculative decoding. This is an infrastructure-heavy model intended for multi-GPU inference.

Those numbers help explain the design, but they do not determine whether Hy4 is right for a specific workflow. Use the MyClaw AI model directory to compare currently listed alternatives by task fit, latency, provider availability, and total cost—not parameter count alone.

What the 1M Context Window Changes—and What It Does Not

The 1M-token context window gives Hy4 room for large repositories, document collections, extensive logs, or long-running agent sessions. It does not guarantee perfect recall, and more context increases input cost. Test whether the model finds the right evidence, follows buried instructions, and maintains decisions across tool calls. A smaller, cleaner context can outperform an unfiltered dump.

🚀 Hy4 preview is here. 770B, 49B active, 1M context. Built for  productivity. Open source frontier. Consistent affordable price. Use it.  Tell us what breaks. More on Hy blog:https://t.co/rbl1IWRk3C  HuggingFace:https://t.co/mE9wevH5XR Github:https ...

What Hy4 Preview Is Built to Do

Long-Horizon Coding and Tool-Using Agents

Tencent positions software engineering as a central Hy4 workload, emphasizing planning, debugging, and verification across extended development tasks. Available serving implementations also support tool calling, allowing an agent to inspect files, run tests, interpret failures, and revise its work without losing the original goal.

Documents, Analysis, Games, and Scientific Work

Tencent also highlights office and analysis tasks that combine messy files into documents, spreadsheets, presentations, equations, or financial models. Game-development examples focus on turning short instructions into playable prototypes and continuing the build over multiple turns.

Scientific work is another stated focus, covering AI research, molecular dynamics, condensed-matter physics, and mathematics. For financial, scientific, or security decisions, still require traceable sources, reproducible calculations, and human review.

How Strong Is Hy4 Preview?

Official Benchmarks and Tencent's Expert Blind Test

Tencent reports broad gains over Hy3 across its benchmark suite. More useful for agent work is its internal blind comparison: 163 Tencent experts rated outputs from 203 engineering tasks. Hy4 preview averaged 2.99, compared with 2.92 for GLM 5.3 and 2.94 for Kimi K3.

Tencent internal comparisonGLM 5.3Kimi K3
Hy4 average vs. comparison model2.99 vs. 2.922.99 vs. 2.94
Hy4 win rate46.8%51.2%
Tie rate12.8%7.9%
Hy4 loss rate40.4%40.9%

The margins are narrow; Hy4 did not dominate either model. As production evidence accumulates, check whether it appears in the OpenClaw AI model benchmark. Until then, prioritize tasks drawn from your own work.

Why Release-Day Scores Are a Starting Point, Not a Verdict

Benchmarks compress a model into one number while hiding prompt design, tool configuration, inference settings, and failure severity. Tencent's blind test is stronger evidence than an anecdotal demo, but it is still an internal evaluation built around Tencent's task mix.

Hy4 should be tested against the exact alternatives available in your stack. Kimi K3, for example, also offers a 1M context window and targets long-horizon coding and agentic knowledge work, but it has a different architecture, feature set, pricing model, and provider path. The practical winner depends on the task.

Try Hy4 Preview in MyClaw

Test Hy4 Preview on your own coding, long-context, and agent tasks—no setup required.

Chat with Hy4 Preview

Hy4 Preview Pricing and Access

API Access Through TokenHub and OpenRouter

At launch, Hy4 preview is available through Tencent Cloud TokenHub and OpenRouter. The OpenRouter model ID is tencent/hy4-preview, with support listed for tool calling and structured outputs.

On August 28, 2026, OpenRouter listed prices of $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache-read tokens. Treat these as a dated snapshot. Provider rates, routing, availability, and promotional terms can change, and an agent may resend substantial context across multiple steps. Estimate the cost of a complete workflow rather than multiplying the price by one prompt.

Self-Hosting the BF16 or FP8 Weights

Tencent released Hy4 preview and Hy4 preview-FP8 model weights under Apache 2.0. The official deployment guidance covers vLLM and SGLang, exposes an OpenAI-compatible API, and includes parsers for reasoning and tool calls.

Open weights do not mean easy local hosting. A 770B backbone requires substantial multi-GPU capacity even at lower precision. Self-hosting fits cases where control, data placement, or sustained utilization justifies the operational work; otherwise, start with a hosted API.

Hy4 Preview Limitations to Test Before Production

tencent/Hy4-preview · Hugging Face

Long Reasoning and Excessive Self-Verification

Tencent explicitly identifies two preview-stage problems: Hy4 may spend longer than necessary reasoning through complex tasks, and it can over-verify its own work. Both can increase latency and token consumption. They can also produce a frustrating pattern in which an agent repeatedly checks a correct result instead of finishing.

Test completion time alongside answer quality. Use deep reasoning only where the task warrants it, define an acceptance test, and tell the agent when to stop. For routine transformations or fast interactions, a smaller or non-reasoning model may deliver a better experience.

Preview-Stage Reliability, Latency, and Version Risk

Release-day availability is not long-term reliability evidence. Provider capacity is still settling, independent tests are limited, and the model may change quickly as Tencent collects feedback. A preview update could improve quality while also altering behavior that a production prompt depends on.

Before adoption, test tool-call validity, structured-output compliance, context retention, timeout behavior, and recovery after a failed tool. Confirm modality support in the provider endpoint you actually use rather than assuming every deployment exposes the same features. Pin configurations where possible and retain a fallback model.

Should You Use Hy4 Preview?

Strong Fits: Coding Agents, Long Documents, and Multi-Step Analysis

Hy4 is most compelling when a task needs both extended context and sustained execution. Good pilots include repository-wide debugging, migration planning, document reconciliation, financial-model review, and research workflows that must turn many files into a structured artifact.

It is also a natural candidate for hosted coding-agent workflows, where the model can inspect a repository, run terminal tools, verify a change, and continue working while the local computer is offline. In these scenarios, judge the complete loop: correct changes, passing tests, useful explanations, elapsed time, and total tokens.

Situations Where a Mature or Faster Model May Be Safer

Choose a more established model when predictable latency, stable behavior, or extensive independent validation matters more than frontier open-weight performance. Simple classification, short-form chat, and high-volume transformations rarely need a 1M context window or prolonged reasoning.

Hy4 is also a poor production choice if the team cannot run a controlled evaluation. A benchmark lead of a few hundredths does not offset unreliable tool calls, unexpected costs, or slow completion on the task that actually matters.

Test Hy4 Preview in a Private, Always-On OpenClaw Workspace

MyClaw provides managed OpenClaw hosting for a private, always-on agent environment. It can serve as a controlled Hy4 test workspace once a compatible provider route is confirmed, without requiring you to operate the agent server yourself.

1. Launch a Managed OpenClaw Workspace

Choose a workspace from the current MyClaw pricing plans based on the tools, files, and concurrent work in your test—not the size of Hy4's remotely hosted weights.

Chat with Hy4 Preview

2. Confirm a Supported Hy4 Provider Route

Check the model and provider options available to your MyClaw account. If a compatible route exposes Hy4, connect the required key and use the exact model ID shown by that provider. Do not assume MyClaw availability from an OpenRouter listing alone.

3. Run Your Own Task Set and Keep a Fallback

After the connection passes a test message and tool call, run representative coding or long-context tasks. Record quality, elapsed time, and token cost, then keep a proven fallback for failures or latency-sensitive work.

The Practical Verdict on Hy4 Preview

Hy4 preview is a credible new open-weight option for long-context, tool-driven work. Its architecture, price, and early engineering results justify a serious pilot. Its preview status, narrow benchmark margins, and official latency warnings rule out blind adoption. Run it on your hardest representative tasks, compare the entire workflow, and promote it only when the operational evidence matches the launch promise.

Ready to Try Hy4 Preview?

Start a chat and evaluate its quality, speed, and cost on work that matters to you.

Chat with Hy4 Preview
Hy4 Preview: Specs, Benchmarks, Pricing, and Who Should Use It | MyClaw.ai