Pull to refresh
Logo
What Changed
Following Why Sign Up
Tencent releases Hy4 with vision and computer control

Tencent releases Hy4 with vision and computer control

New Capabilities

Tencent's 770-billion-parameter model adds image understanding and software control, but the weights are still unpublished

2 days ago: Tencent reports finance-professional scores for Hy4

Overview

Updated Yesterday

Tencent released the full version of Hy4 on November 4, adding image understanding and the ability to operate software. The preview had 770 billion total parameters, 49 billion active per token; Tencent did not disclose a revised count for the full release. Hy4 is sold through Tencent Cloud and OpenRouter and runs inside Tencent's coding, office and chat products.

The weights are still unpublished. As of November 9, Tencent's launch post said they and a system card would follow, with no download link. On November 5 Tencent said it had expanded its Hy4 inference cluster tenfold after WorkBuddy users hit free quotas, and free access in WorkBuddy and CodeBuddy ends November 15.

Why it matters

A model that reads charts and clicks through software costs $0.83 per million input tokens, and Tencent still has not released the weights.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

770B
Total parameters (preview)
The 770 billion total and 49 billion active figures are from the August preview; Tencent did not disclose a revised count for the full release.
$0.834
API price per million input tokens
Unchanged from the preview; Tencent's release table lists $0.042 for cached input and $2.501 for output.
$2.501
API price per million output tokens
Output costs roughly three times input. Tencent quotes CNY 6 input and CNY 18 output per million tokens.
7.2%
Hallucination rate on Tencent's internal evaluations
Down from 12.95% on the preview. On internal search evaluations the rate fell from 15.76% to 5.45%.
19.1%
Game development improvement over preview
Tencent's internal GameDev benchmark; game logic improved 9.8%.
72%
Finance evaluators from outside Tencent
153 of the 213 finance professionals who scored Hy4 on 579 tasks came from external institutions. Hy4 averaged 3.04, against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for Kimi K3.1 Pro and GLM-5.3.
10x
Hy4 inference capacity added
Tencent said on November 5 it expanded its Hy4 inference cluster tenfold after WorkBuddy users hit free quotas on launch day. It gave no absolute figures.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

Organizations Involved

Tencent
Tencent
Technology company
Developer and distributor of Hy4

Shenzhen-based internet company that builds the Hunyuan model family and ships it through its own apps and cloud.
Tencent Hunyuan
Tencent Hunyuan
AI research and product unit
Built and trained Hy4

Tencent's model team, which trains the Hy series and publishes its research notes.
OpenRouter
OpenRouter
Model hosting marketplace
Third-party channel for Hy4 access

A marketplace that routes developer requests to many AI models behind one API.
VG
Vercel AI Gateway
Developer infrastructure service
Lists Hy4 alongside other models

A routing layer that lets web developers swap AI models without rewriting their code.

Timeline

August 2026 November 2026

5 events Latest: 2 days ago
Tap a bar to jump to that date

  1. Tencent reports finance-professional scores for Hy4

    Latest Evaluation

    213 finance professionals scored Hy4 on 579 tasks, 153 of them from outside Tencent. Hy4 averaged 3.04, against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for Kimi K3.1 Pro and GLM-5.3; Tencent organized the evaluation and the results were not independently verified.

  2. Tencent expands Hy4 inference capacity tenfold

    Infrastructure

    Tencent said it added ten times the inference capacity after WorkBuddy users hit free quotas on launch day. It gave no absolute figures and warned peak-hour limits may persist.

  3. Full Hy4 released with vision and computer use

    Release

    Tencent adds image understanding, command-line and graphical interface control, and long-horizon task training.

  4. Hy4 goes live on Tencent Cloud and OpenRouter

    Distribution

    Pricing held flat versus the preview at $0.834 input and $2.501 output per million tokens.

  5. Hy4 preview ships with weights

    Release

    Tencent open-sourced a preview version and named over-reasoning and over-verification as known flaws.

Scenarios

1

Tencent publishes Hy4 weights and system card

Likely Resolves by Q1 2027

Discussed by: Tencent's own release post, which says weights and a system card will follow

Tencent has said it will open-source the weights and pair them with a system card covering safety evaluations. The preview was open-sourced in August, and the company has lined up the vLLM and SGLang inference projects to support serving. As of November 9 the launch post still described the files as coming in the coming days, with no download link. A downloadable 770-billion-parameter model would let cloud providers and enterprises run Hy4 on their own hardware instead of paying per token. Tencent has not given a date.

2

Hy4 Flash launches at a lower price than Hy4

Likely Resolves by Q2 2027

Discussed by: Tencent's release post, which says a flash model is coming soon with better cost-efficiency

Tencent said it plans a Hy4 Flash variant with faster inference and lower cost. That would mirror the tiering most labs use: a large model for hard tasks, a cheap one for volume. If it ships, the practical effect is that high-volume work like document parsing and routine coding help gets cheaper, while Hy4 stays the option for long multi-step jobs.

3

Hy4 becomes a top-10 model by token volume on OpenRouter

Possible Resolves by Feb 28, 2027

Discussed by: Developer routing data published by OpenRouter

Hy4's availability on OpenRouter and Vercel's AI Gateway puts it in front of developers who never touch Tencent Cloud. Price is the lever: $0.834 per million input tokens is cheap for a model with vision. Whether usage follows depends on whether Hy4 holds up on coding and agent tasks against cheaper Chinese rivals and US mid-tier models.

4

Tencent keeps the free Hy4 quota running past November 15

Possible Resolves by Nov 16, 2026

Discussed by: EastFrontier, which notes Tencent has not said whether the raised quotas survive the deadline

Tencent raised free usage allowances for Hy4 and extended free access to Hy3 on WorkBuddy until 23:59 on November 15. It has not said whether the raised quotas continue after that date. If they do, it signals Tencent is willing to absorb the serving cost to hold WorkBuddy users; if they do not, usage data after November 15 will show how much of the demand was price-driven.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

July 2023

Meta releases Llama 2 weights (July 2023)

Meta published the weights for Llama 2 and allowed commercial use, a break from the narrower research licenses that had come before. Microsoft agreed to distribute it through Azure.

Then

Startups built products on Llama 2 immediately, and hosting services sprang up to run it.

Now

It set the template for open-weight releases that Chinese labs later followed and, in several cases, outran.

Why this matters now

Hy4 follows the same playbook: publish weights, line up inference tools, and let outside companies distribute it.

December 2024

DeepSeek-V3 ships (December 2024)

DeepSeek released a mixture-of-experts model with 671 billion total parameters but only 37 billion active per token, which kept serving costs low. It matched or beat larger US models on several coding and math tests.

Then

It became the default cheap model for developers experimenting with open weights.

Now

Chinese labs began competing openly on price-per-token rather than only on benchmark scores.

Why this matters now

Hy4's architecture and $0.834 input price put it in the same contest: large total parameters, small active slice, aggressive pricing.

January 2025

DeepSeek-R1 release (January 2025)

DeepSeek published an open reasoning model and its app topped download charts. Investors read it as evidence that Chinese labs could match US frontier work at far lower cost, and AI chip stocks fell sharply.

Then

US labs publicly disputed the cost claims and accelerated their own release schedules.

Now

Open-weight Chinese models became a permanent part of the global market rather than a curiosity.

Why this matters now

Tencent's claim to be in the top tier of open models is the next step in that arc, now with vision and software control attached.

Sources

(16)
+10
Tencent Hy TechNode Unite.AI South China Morning Post 科创板日报 (CLS) Vercel Nikkei Asia Tencent PR Newswire IAMAG ClaudeGPT Si TNGlobal EastFrontier AI Agents Info Dataloco Caixin Global