Tencent Hy4 flash cuts token prices, but task savings need a separate calculation

Text
iHOLYSHITVISUAL / TECHModel releasesiHOLYSHIT / ILLUSTRATION
iHolyShit illustration

Tencent’s Hy4 flash combines a million-token context window with lower API prices. Its reported twentyfold task-cost advantage comes from an internal office-work evaluation. Temporary discounts are separate from list prices, and open-source release remains a plan rather than an announced weight release.

Tencent launched Hy4 flash on November 16 with list prices of RMB 0.5 per million input tokens, RMB 2 for output and RMB 0.02 for cache hits. Tencent reports one-twentieth of Hy4’s per-task cost on its internal WorkIndex office-task evaluation, not a universal saving.

Cheaper tokens and cheaper completed work are different measures. Retries, tool calls and task difficulty affect the latter. For a workflow handling many documents, reliable completion determines whether a lower rate translates into a lower bill.

The official catalog lists image and text input, text output, tool calling, structured output and caching. It specifies a 1,024K context window, up to 960K input and 64K output tokens. These limits describe different things: accommodating more source material does not imply an equally long answer or perfect use of every detail.

Tencent lists 224 billion total parameters, with 12 billion active. Total and active parameters describe different quantities. Neither the model’s full size nor its activation ratio directly determines a customer’s bill. Model scale, service pricing and the resources consumed by a completed task need separate comparisons.

Before November 30, WorkBuddy offers limited-time free use and TokenHub offers half-price access. The catalog shows discounted rates of RMB 0.25 for input, RMB 1 for output and RMB 0.01 for cache hits per million tokens. Tencent says it plans to open-source the model, not that weights have already been released. An API promotion is not a self-hosting cost estimate.

iHolyshit review

I would treat this as a candidate for repetitive work, not use one cost ratio to dismiss a flagship. Run the same document workload and compare errors, rework and the total bill. Making retries affordable is progress; making mistakes acceptable is not.

Share this article

What do you think of this story?

Anonymous feedback, up to 5 reactions per article.

Reader comments

Comment without signing in, or sign in to display your member name.

Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.

0/500

No comments yet. Be the first to share.

Sources