Tencent Hy4 Preview makes long context more practical. The hard part is keeping the decisive evidence

Tencent’s Hy4 Preview is an attempt to make substantial working context useful for coding, office tasks and research. Released on 28 August, it is offered through open weights and hosted routes including Tencent Cloud TokenHub, OpenRouter and Tencent applications. That provides several ways to investigate the same model family—but not necessarily identical runtimes, limits or bills.[1]

Its central opportunity can be expressed as a concrete task. Give an assistant a large project archive and ask whether a proposed change is safe. The important answer may depend on one obsolete-looking file, a later exception or an interaction between two distant documents. A larger reading window helps only if those relationships survive the model’s selection of what matters.

A cheaper way to decide where to look

Hy4’s implementation uses compressed attention representations and dynamic sparse attention. The model card describes IndexCache, which reuses attention-selection information across layers rather than recomputing every selection independently. The intended saving is in the machinery deciding which earlier information deserves attention.[2][3]

The IndexCache research investigates the similarity of selected positions across adjacent layers. It develops both training-free and training-aware approaches to reducing indexer work through reuse. The paper is evidence for a computational technique; it is not a guarantee that a particular customer’s most important sentence will always be selected.[4]

An analogy helps separate efficiency from correctness. A researcher can use a colleague’s list of relevant pages instead of rebuilding the index from scratch. That saves effort when the list remains appropriate. But the final judgement still depends on reading the right evidence, interpreting it correctly and noticing when a supposedly irrelevant exception changes the case. Reusing a good index is not the same as proving the conclusion.

A million-token contract can fail on one sentence

Here is an original evaluation design for a procurement team. Construct a fictional agreement set containing a main contract, a later amendment and a small schedule. The main contract permits a service; the amendment narrows that permission; the schedule exempts one category. Ask for the permitted action, not a summary of each document. The answer must identify the applicable version and connect the exception to the specific case.

Then change only one controlling sentence. Keep filenames and document order stable. A system that produces almost the same confident answer in both runs may be relying on the broad theme rather than the decisive rule. A successful test requires the conclusion to change for the right reason, with the changed source cited. This is more revealing than asking the model to repeat a hidden phrase.

How much weight to put on Tencent’s expert score

Tencent reports an internal blind evaluation involving 163 experts and 203 engineering tasks. Hy4 averaged 2.99 out of four, compared with 2.94 for Kimi K3 and 2.92 for GLM-5.3. That is a company-run assessment, and the announcement does not supply uncertainty intervals that would establish how robust those small differences are.[1]

The strongest use of that disclosure is as a lead for further investigation, not a universal ranking. Were the tasks close to the buyer’s work? Did the reviewers see final artefacts or only conversations? How much intervention was allowed? A score can be informative while leaving these questions unresolved. A 0.05-point lead on a four-point scale cannot by itself determine a procurement decision.

There is also a practical difference between a hosted trial and a self-hosted rollout. The first can establish task fit quickly. The second adds decisions about weight precision, serving implementation, concurrency, logging and access control. A company may value control of the deployment, but that control comes with operational work. The model’s active-parameter figure does not eliminate the rest of the system.

Make the evidence portable

For a pilot, we would store the selected model, runtime, prompt, source-file hashes, context limits and final artefact together. A reviewer should be able to open the cited passage without repeating the entire conversation. That record also makes a later upgrade meaningful: the same task can be replayed under a new checkpoint without quietly changing the evidence at the same time.

The broader significance of Hy4 is that it expands the set of large, inspectable models available for this kind of work. Our preferred success criterion is correspondingly demanding: not the number of pages accepted, but the proportion of consequential decisions whose evidence chain survives scrutiny. Sparse attention can make the work affordable. The evidence chain determines whether it is worth doing.

Report an error or suggest a correction