Hy4 preview

tencent/hy4-preview-20260827

TencentChinaOpen weights

The most-used model on the router over the last 7 days, with 15.5% of tokens. Weekly share peaked at 12.7% in the week of Aug 31, 2026, with 18T tokens processed since Aug 24, 2026.

Launch: Aug 28, 2026
Context: 1M tokens
Modality: text → text
Reasoning: yes
License: Open weights
Input: $0.834 / 1M
Output: $2.50 / 1M

Share, last 7 days

15.5%

1st of 61 · Sep 4, 2026 to Sep 10, 2026

Peak weekly share

12.7%

week of Aug 31, 2026

Cumulative tokens

18T

since the week of Aug 24, 2026

Weeks in the top 10

of 2 weeks with volume

Weekly token share

Share of the router's weekly volume, since the first week with volume

See the numbers

Week	Share
Aug 31, 2026	12.708%
Aug 24, 2026	2.713%

Debuted in the week of Aug 24, 2026 and peaked at 12.7% 1 week later. The latest full week is the peak so far. Over the last 7 days, which already reach into the current week, it stands at 15.5%, above the weekly peak.

What this model is used for new

Tasks where the model ranks among the leaders, and its share of each

File read and write26.7%5.9%

Workflow execution22.8%24.6%

Research and reports22.8%1.2%

Conversation21.6%3.1%

Multi-step planning10.2%6.6%

Finance and trading10.2%0.40%

Web search8.2%0.80%

Frontend and UI5.8%2.9%

SQL and databases5.6%0.50%

Shell execution5.0%2.6%

Tool calling4.8%1.2%

Code generation4.7%13.1%

Plus 5 more tasks with a smaller share.

Ranks among the leaders in 17 of the 29 classified tasks. Its largest share is in file read and write, 26.7% of the task's tokens; in the heaviest task where it appears, workflow execution (24.6% of classified volume), it holds 22.8%.

Snapshot of the rolling 7-day window through Sep 10, 2026. Share: the model's tokens in the task over the task's total tokens. Weight: the task's share of classified volume.

Where to run it

Endpoints in the source's catalog, from cheapest to most expensive on input

Provider	Input	Output	Context	Quantization	Zero retention
Tencent	$0.834	$2.50	1M	fp8	yes

1 endpoint across 1 provider. 1 offers zero data retention.

Prices in US dollars per million tokens, from the latest archived endpoint catalog.

Benchmarks

Artificial Analysis indexes and benchmarks run by the source, with cost per task

No Artificial Analysis index and no source-run benchmark for this model. A missing benchmark is not a low score.

Against the week's largest models

The five largest proprietary and five largest open-weights models over the last 7 days

Model	Lab	7-day share	Blended price	Context	Intelligence	Weeks in top 10
Proprietary
GPT-5.6 LunaOpenAI	OpenAI	11.5%	$0.45	1M	37.5	6
Gemini 3.8 FlashGoogle	Google	2.1%	$1.50	1M	41.2	0
Muse Spark 1.3 ContributorMeta	Meta	1.5%	$0.125	1M	—	0
Claude Opus 5Anthropic	Anthropic	1.4%	$10.00	1M	50.7	1
GPT-5.6 SolOpenAI	OpenAI	1.4%	$4.00	1M	47.1	0
Open weights
Hy4 previewthis pageTencent	Tencent	15.5%	$1.25	1M	—	1
DeepSeek V4 Flash 0731DeepSeek	DeepSeek	10.0%	$0.094	1.31M	34.5	5
GLM 5.3 FlashZ.ai (GLM)	Z.ai (GLM)	10.0%	$0.237	1.31M	41.9	2
MiMo-V2.5Xiaomi	Xiaomi	4.2%	$0.175	1M	22.3	15
DeepSeek V4 Flash 0423DeepSeek	DeepSeek	3.9%	$0.111	1M	24.8	19

Share over the last 7 days (Sep 4, 2026 to Sep 10, 2026). Blended price in US dollars per million tokens, weighted 75% input and 25% output. A dash means the number is not published, not zero.

What this page doesn't say. Share here is a slice of one router's traffic, where the deciding factor is usually price per token. It is not market, revenue or user share, and it does not tell you which model fits your workload.