Is Tencent Hy3 available with Zero Data Retention on Elosia?

Tencent Hy3 runs on Elosia's standard backend; ZDR is available on many other models in the catalog.

How much does Tencent Hy3 cost on Elosia?

Tencent Hy3 is billed by usage at $0.0825 per 1M input tokens and $0.33 per 1M output tokens, within your Elosia plan.

What is Tencent Hy3's context window?

Tencent Hy3 supports a context window of up to 262,144 tokens.

Is Tencent Hy3 hosted in the EU?

Tencent Hy3 is not available on Elosia's EU-hosted (EdenAI) backend; it runs on the default backend.

When should you choose Tencent Hy3?

Tencent Hy3 is a balanced model: a strong quality/cost trade-off for everyday production work.

Tencent Hy3

TencentBalanced

ThinkingTool UseStructured Output

About this model

Hy3 is a 295B-parameter Mixture-of-Experts model from Tencent (21B active, 192 experts with top-8 routing) built for reasoning, agentic workflows, and real-world production use. It supports a configurable reasoning effort:...

Performance Tier

Balanced

Tencent Hy3 is a balanced model from Tencent : strong performance at a reasonable price.

Strong cost-performance ratio. Reliable for most professional use cases without premium pricing.

Pricing

This model is included in Elosia plans

Eco

Minimal cost. Ideal for very high volume or simple tasks.

Type	per 1M tokens
Input (prompt)	$0.083
Output (completion)	$0.330
Cache read	$0.021

Capabilities

Context Length262K

Max Output Tokens128K

TokenizerOther

Inputtext

Outputtext

Release DateJuly 6, 2026

Benchmarks

General Intelligence

MMLU

Not reported

GPQA Diamond

90.4%

Mathematics

MATH-500

Not reported

Programming

HumanEval

Not reported

SWE-bench Verified

78%

SWE-bench Multilingual

75.8%

Reasoning

Humanity's Last Exam

37%

Agentic

SWE-bench Pro

57.9%

Terminal-Bench 2.1

71.7%

Where does Tencent Hy3 stand?

Compare its performance index against every other model.

View the performance leaderboard

Recommended Use Cases

CodingResearchAnalysisGeneral Chat

Strengths

Sparse mixture of experts: 192 experts with top-8 routing over 80 layers, so each token engages only a small share of the network
Agentic search is an explicit evaluation axis, with multi-step browsing and deep-search suites reported next to the coding and reasoning ones
Genuine openness: Apache 2.0 weights plus a published FP8 build, documented vLLM and SGLang deployment, and only 21B active parameters out of 295B, which keeps the per-token compute low
Tool calling through a dedicated hy_v3 parser, on a 262K-token context window with up to 128K of output

Limitations

Weak factual coverage: faced with a question it does not master, it invents more often than it abstains, which tempers the anti-hallucination claim of the model card
Competition mathematics is the soft axis, in Tencent's own table, well behind its scientific reasoning. Text only, no multimodality