Tencent opens Hy4 flash weights. What capacity would self-hosting take?

The instruct, FP8 and pretrained releases make self-hosting an option. Capacity estimates start with the exact checkpoint and continue through real request sizes, concurrency and operating costs.

Article published: 2026-11-20Mokyun Editorial4 min read

Public developments and original analysis · Announcement date: · Sources

Tencent introduced Hy4 flash and its hosted access on November 16 and announced downloadable weights on November 18. Its repository lists an instruction-tuned model, an FP8 version of that model and a pretrained base, with links to Hugging Face, ModelScope and GitCode. The model accepts text and images and generates text. The documentation specifies a context window of roughly one million tokens. Companies processing their own files can now choose a runtime and a checkpoint for internal deployment.

AI-generated conceptual illustration. A lit group of tiles within a complete ceramic relief represents a subset of model parameters taking part in a computation.
AI-generated conceptual illustration. A lit group of tiles within a complete ceramic relief represents a subset of model parameters taking part in a computation.

The roughly 12 billion active parameters describe how much of the model participates in computation for a token. Tencent's launch announcement describes 224 billion total parameters, while the repository describes approximately 220 billion and lists the backbone, Engram, vision encoder and draft decoder separately. For capacity planning, start with the files for the exact checkpoint and the selected engine's loading behavior. Different requests can route to different experts, and the service has to make their weights available.

Sizing the service from its files

Add up the checkpoint files to estimate storage, distribution and update costs. Account separately for how the engine places weights, keeps intermediate state and schedules concurrent requests. Tencent documents an FP8 configuration with multiple accelerators. Use it to identify software and hardware assumptions, then check whether the available equipment can handle the intended workload, especially requests that combine images and long documents.

Consider a hypothetical parts supplier extracting model identifiers from nameplate photos and preparing records for engineers to check against a manual. The image step has to preserve distinctions as small as a single character. The text step has to cite the correct edition of the manual. Start with the photo sizes and image quality the team receives. Then test the wait when several pages are attached to the same request. Retrieving the relevant sections by product model can reduce the volume and give the engineer a more direct route to the source page.

A batch arriving at the start of a shift may affect other requests. A few long jobs can increase queueing and memory use, so measure peak memory and the long end of the latency distribution with short and long jobs running together. A successful run on one clear photo establishes that the system can load and produce its expected format. Capacity estimates also need unclear images, busy periods and retries when the wrong document has been supplied. Separate queues can give interactive review a predictable response while batch extraction uses spare capacity.

Choosing a checkpoint and keeping it usable

The instruction-tuned releases are the candidates the documentation uses for conversations and instructions. FP8 provides a way to compare memory and output quality on a supported engine and hardware. The pretrained base is intended for work that involves further training, which adds responsibility for training samples, evaluation and release management. A team with a stable extraction workflow can compare the two instruct checkpoints on the same photos and reference answers. Counts of character errors, missing fields and minutes of correction time will show whether any capacity gains help the actual job.

Self-hosting lets a team pin engine and model versions and gives it responsibility for upgrades and recovery. A deployable package should identify the model revision and the quantization, and have a tested path to restoring a previous version or rebuilding after a failed container or cache. Access for external customers also calls for a reading of the Tencent Hy License 1.0. Section 10 sets out the definition of a model service, conditions under which providers at certain scales need a separate license, and an exception for solely internal use when the relevant materials and outputs are kept internal. The way a product is actually delivered determines which conditions to review.

Start with the part of delivery that limits the current workflow. If hosted access meets the data requirements and demand varies with projects, metered use can remain practical while the same evaluation set is kept for future comparisons. Stable traffic, explicit data boundaries and staff who can operate the service make a self-hosted option worth measuring. Compare time to complete a usable work order, human correction effort and total monthly cost to choose a checkpoint and deployment route.

Sources

  1. Tencent Hy4 flash launch announcement

    Source published: · Accessed:

  2. Tencent Hy4 flash model and deployment documentation

    Accessed:

  3. Tencent Hy License 1.0

    Accessed:

  4. Tencent Hy4 flash download and model card

    Accessed:

Mokyun Editorial · neo@mokyun.com