Hy4 API

Hy4: Tencent's 770B-class MoE flagship

Hy4 is the final release of the Tencent Hy team's flagship Mixture-of-Experts model, published on November 4, 2026. Its preview, released August 28, 2026, shipped as a sparse MoE with 770B total parameters, 49B active parameters per token, and a 1M-token context window. Tencent states the final Hy4 substantially scales up both pre-training and post-training relative to that preview.

Hy4 release timeline

The Hy family shipped in two steps. Knowing which checkpoint you are calling matters more than the model name alone, because pricing, quality and rate limits can differ between them.

Release dates from the Tencent Hy research page. Verification of enabled versions on TokenRa accounts should always come from the live console.

Hy4 vs Hy4 Preview: what changed

Tencent describes the final Hy4 as a substantial scaling-up of both pre-training and post-training rather than a patch on the preview. The practical differences below are the ones worth re-testing.

Where a cell says verify, the value is channel- and account-specific; read it from the model directory or console.

Where Hy4 API fits

Tencent built preview-era messaging around agentic work, coding, and productivity scenarios. Test the model on the shapes of input your application actually receives, not on short demo prompts.

• Coding agents that read a repository, plan a change, call tools, and retry across multiple steps.
• Office and finance workflows that extract fields from long documents and must cite the source clause.
• Frontend and game prototyping where the model iterates on rendered output rather than a single textual answer.
• Long-running assistants that manage history, tools, and user state across many turns.
• Research workflows that summarize many supplied sources and need traceable conclusions.

Use the OpenAI SDK to call Hy4

Limit each request to what the task needs: a clear system instruction, the smallest sufficient context, and an explicit output format. Retries, truncation and citation checks belong in your application, not in the prompt alone.

from openai import OpenAI

client = OpenAI(
    base_url="https://tokenra.io/v1",
    api_key="YOUR_API_KEY",
)

response = client.chat.completions.create(
    model="hy4",  # use the exact identifier from the console
    messages=[
        {"role": "system", "content": "You are a precise coding assistant. Cite the files you use."},
        {"role": "user", "content": "Review the pasted module and list concrete defects."},
    ],
)
print(response.choices[0].message.content)

Hy4 API pricing and model ID

Hy4 is billed per 1M tokens, with separate input, output and cached-read rates. Pricing on TokenRa is set per channel, so confirm the current rate and model identifier in the console before integrating.

Model ID: hy4. Replace with the identifier listed in the console once the channel is enabled.

Production checklist

• Re-run your preview-era evaluations; a later training run is not a drop-in swap for prompts tuned on the old checkpoint.
• Keep API keys server-side and separate credentials by environment or workload.
• Set explicit timeouts, bounded retries, and request tracing before routing user traffic.
• Cap context growth: trim stale turns, summarize history, and monitor usage as conversations lengthen.
• Budget for the final release rate rather than the preview or free promotion.
• Review provider terms, privacy requirements, and output rights for your use case.

Hy4 API FAQ

Continue exploring