Releases

Tencent Hy4 Flash

MiniRouter and Tencent Hy4 Flash by Tencent, separated by a silver fold of light on black.

Tencent's smaller Hy4 model for frequent coding, document and agent tasks, with 12B active parameters, a 1M-token context and optional reasoning.

Tencent Hy4 Flash is available on MiniRouter as . The model page has current pricing, provider routes and API setup.

View Tencent Hy4 Flash

Less active compute for everyday work

Tencent introduced Hy4 Flash on November 16, 2026 as a smaller sibling of Hy4. Its design combines hybrid attention, LatentMoE and Engram memory, with 12 billion active parameters out of 224 billion total.

The useful distinction is the workload: frequent document processing and coding assistance, with the larger Hy4 reserved for complex, long-running software engineering. A smaller model is worth evaluating when repeated drafts, summaries and proposed edits make up much of your traffic.

Tencent launch report, November 16, 2026. Architecture and positioning are publisher-reported.

Make each step easy to verify

For a change proposal, gather the relevant specification and code, ask for a draft, then check every proposed action. The same pattern works for turning several text documents into a sourced brief.

A document-to-draft workflow
  1. Prepare the text

    Collect the relevant passages, code and task instructions. Give each source a label so the answer can point back to it.

  2. Draft with Flash

    Ask for a focused summary, a proposed edit or an explanation. Keep each step small enough to check independently.

  3. Check and escalate

    Compare claims with the sources and run tests for code. Move unresolved, long-running engineering work to a stronger model.

Start with extracted text. MiniRouter currently accepts text input for Hy4 Flash; image inputs and schema-constrained outputs are not available on this route. Your application supplies any tools and validates their results.

Room for sources, with a separate output limit

The gateway lists a 1,024,000-token context and a 64,000-token maximum output. Those limits describe capacity, not a target request size. Send the material needed for the question and ask for the smallest useful deliverable.

View endpoint specifications
Hy4 Flash endpoint limits and reasoning options
SettingPublished value
Context window1,024,000 tokens
Maximum output64,000 tokens
Reasoning effortnone or high
Vercel endpoint specifications, read November 16, 2026.

Reasoning is optional. Explicit reasoning settings can reserve the full output ceiling before the request; the final charge follows actual usage.

Judge the finished task

Tencent publishes internal WorkIndex results and a WorkBuddy user study. These are useful launch evidence, but a task cost depends on the prompts, tools and checks around the model. They do not establish an independent ranking or MiniRouter response times.

Build a small evaluation set from your own work. Count accepted answers, retries and total tokens, and check whether errors require a more capable model. For code, use the same tests for both candidates; for document work, check every citation against the supplied text.

Tencent evaluation report. Publisher-run studies, November 2026.

Check the current price before a larger run

Vercel announced introductory pricing through November 29, 2026. Use the model page for current MiniRouter rates; this article does not lock in a launch price.

Gateway announcement · Tencent Cloud pricing

Current prices and API setup →

Keep exploring

Compare published results in the model rankings and coding benchmarks, or read more model releases.

For MiniRouter product updates, visit the changelog.

Sources

  1. Tencent Hy4 Flash announcement

    Published Nov 16, 2026

    Read at Tencent ↗
  2. Hy4 Flash on Vercel AI Gateway

    Published Nov 16, 2026

    Read at Vercel ↗
  3. Hy4 Flash endpoint specifications

    Published Nov 16, 2026

    Read at Vercel ↗
  4. Tencent Cloud TokenHub pricing

    Published Nov 16, 2026

    Read at Tencent Cloud ↗

Editorial reference revised 2026-11-16. Catalog status comes from the generated routing snapshot.

Every model, one key.