Open SourceTencentReleased 2026-11

Hy4

Tencent's flagship model, the first Hy model with vision and computer use

Context window

1M

Input / 1M tokens

$0.834

Output / 1M tokens

$2.501

Provider

Tencent

Tencent Cloud TokenHub list price (Singapore), per 1M tokens: $0.834 input, $2.501 output, $0.042 cached input, the same as Hy4 Preview; Tencent rounds this to about $0.90/$2.70. TokenHub sells Hy4 at 50% off through December 31, 2026, and OpenRouter lists it at the discounted $0.417/$1.25. Max input 960K tokens, max output 64K. Tencent says the weights will follow; as of November 11, 2026 only the Hy4 Preview weights (Apache 2.0) are on Hugging Face. · Data verified 2026-11-11

Hy4 (November 4, 2026) is Tencent's latest flagship large language model and the general release of Hy4 Preview. It is the first Hy model with image understanding and computer use across command-line and graphical interfaces, and Tencent says it was trained on professional software such as Blender and CAD for long, multistep tasks. It has a 1M-token context and costs about $0.83/$2.50 per 1M tokens on Tencent Cloud. Tencent presents it as one of the leading open-source models and says weights and system cards will be released in the coming days; so far only the Hy4 Preview weights are public. A cheaper Hy4 Flash is scheduled for release soon.

How to access it

Tencent Cloud TokenHub API (model ID hy4) through the OpenAI Chat Completions, OpenAI Responses or Anthropic Messages protocols; OpenRouter (tencent/hy4); and Tencent products including WorkBuddy, CodeBuddy, Yuanbao and ima. WorkBuddy and CodeBuddy are free from November 4 to 15, 2026.

Strengths

  • ✓Terminal-Bench 4.0 44.4% vs 42.9% for GLM-5.3 and 40.0% for DeepSeek V4.1 Flash 1017 (Tencent-reported)
  • ✓OSWorld 2.1 78.9% (Tencent-reported)
  • ✓Image understanding and computer use, new in this generation
  • ✓About $0.83/$2.50 per 1M, half off on TokenHub through 2026
  • ✓1M-token context

Best for developers who...

Office and multi-app agent workflows, such as Tencent WorkBuddyFront-end and game development (Unreal Engine, Unity, Godot)Financial research and analysisLow-cost agents that need vision

When to choose it (and when not to)

Reach for Hy4 when...

  • →You want a low-cost model with vision and computer use for multi-app agent tasks
  • →You build on Tencent Cloud or Tencent products such as WorkBuddy and CodeBuddy
  • →You plan to self-host once Tencent publishes the weights

Look elsewhere if...

  • ✕Hard long-horizon coding: it trails Claude Opus 5.5 and Kimi K3.1 Pro by over 20 points on Terminal-Bench 4.0 in Tencent's own chart
  • ✕You need outputs longer than 64K tokens
  • ✕You need the final weights today (only the preview weights are public)

How to use it

  • ›Use function calling and structured output for agent steps; TokenHub lists both for Hy4
  • ›Keep input under 960K tokens, the TokenHub maximum for the 1M window

Quickstart

Python
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["TOKENHUB_API_KEY"],
    base_url="https://tokenhub-intl.tencentcloudmaas.com/v1",
)
response = client.chat.completions.create(
    model="hy4",
    messages=[{"role": "user", "content": "Hello, please introduce yourself."}],
)
print(response.choices[0].message.content)

Tencent Cloud TokenHub is OpenAI-compatible (Chat Completions and Responses) and also accepts the Anthropic Messages protocol. Chinese mainland region: https://tokenhub.tencentcloudmaas.com/v1.

API model id: hy4

Benchmarks

BenchmarkScoreNotes
Terminal-Bench 4.044.4%Tencent-reported (harness not stated); Kimi K3.1 Pro 66.8%, Claude Opus 5.5 66.4%, GLM-5.3 42.9%
OSWorld 2.178.9%Tencent-reported; Claude Opus 5.5 81.8%, Kimi K3.1 Pro 79.6%
AutomationBench v1.0.656.2%Tencent-reported; Kimi K3.1 Pro 65.6%, Claude Opus 5.5 58.8%
FrontierCode 2.031.0%Tencent-reported; Claude Opus 5.5 46.4%, DeepSeek V4.1 Flash 1017 28.5%
Terminal-Bench-Science 0.119.8%Tencent-reported; GPT-6.1 Sol 60.9%, Claude Opus 5.5 56.2%

Source: Tencent - Hy4 launch announcement

Compare Hy4

Compare Hy4 with any other model

Build a comparison →
All model comparisons →

Learn the concepts