Intel / post

Tencent releases open-source Hy3 preview for agents and coding

Source: x.com
Date: Published Apr 26, 2026

Context

In this April 26 post, Tencent attached two numbers to Hy3 preview, its open-source model for coding, search and tasks: a 256K and a "40% inference efficiency gain." The context window is how much text the model can consider in a single request, counted in tokens (word fragments), so 256K tokens lets it work across a large set of code files or several long documents without splitting them up. That figure is a fixed specification, and Tencent's Hy3 preview announcement gives the same limit.

The 40% figure is harder to use. Neither the post's text nor the announcement says what it was measured against or on which workload; the announcement credits it to tuning the model's design and its serving software together, including faster computation and , which stores the model's numbers at lower precision so it needs less memory. The same announcement says that in Tencent's CodeBuddy coding assistant and WorkBuddy work agent, the wait for the first token of a reply fell 54% and total response time fell 47%, again without naming the earlier setup. These are Tencent's measurements in its own products, so they do not show what speedup someone running the on their own hardware should expect.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.

Read the source x.com