Tencent Hunyuan logo
Tencent Hunyuan logo

Tencent Hunyuan

Tencent Hunyuan

MODEL DIRECTORY

MODEL DIRECTORY

Tencent Hunyuan

Tencent Hunyuan

Tencent Hy3 is a fresh 295B-parameter MoE with only 21B active parameters, hybrid fast/slow reasoning, 256K context, and a strong early independent score. The relationship between preview weights and the production Hy3 service still needs verification.

CURRENT MODEL SNAPSHOT

Provider: Tencent Hunyuan
Current anchor: Hy3
Lifecycle: Current
Weights: Mixed; preview weights published, production parity to verify
Reviewed: 21 July 2026

Abstract blue and purple gradient

Hy3 is Tencent’s efficiency-first agent model

Tencent released Hy3 on 6 July 2026. The model uses 295 billion total parameters with roughly 21 billion active per token, a 256,000-token context window, and a hybrid fast/slow reasoning system. The aim is to switch between direct responses and deeper deliberation without requiring a different model family.

That active-parameter ratio is the architectural story. Hy3 tries to deliver frontier-challenger reasoning and tool use while keeping inference work closer to a much smaller dense model. Tencent positions it for coding, research, and product integrations through Tencent Cloud and partner access.

Preview weights are public through Tencent’s Hy3-preview repository. Buyers should not assume that the preview artifact, the 6 July production release, and every hosted endpoint are identical. Weight availability is therefore best described as mixed until the exact current artifact and license are verified.

Early independent evidence is strong, but the release is young

Artificial Analysis reports an Intelligence Index score of 41, around 59 output tokens per second, and a time to first token near 2.5 seconds for the evaluated Hy3 endpoint. The model produced roughly 140 million output tokens across the suite, which suggests that strong task performance may come with verbose reasoning and higher downstream processing cost.

The result is notable because it arrives with only 21B active parameters. It is not yet a mature operational record. Endpoint behavior, quotas, regional availability, documentation, and the parity between preview and production artifacts can change quickly after launch.

  • The independent index supports treating Hy3 as a serious challenger, not a speculative announcement.

  • Measured verbosity means cost must be evaluated per completed outcome, not only per token.

  • A very recent release needs version-pinned regression testing before customer commitments.

The enterprise case: efficient capability with release-management risk

Hy3 is worth testing for Chinese-language and global tool workflows, coding agents, and organizations already operating in Tencent Cloud. Its efficiency could matter at scale. The main caution is the age of the release and incomplete evidence about artifact parity, licensing, and stable enterprise controls across access routes.

How we would evaluate it

Pin the exact endpoint or preview weight checksum and run a multilingual tool workflow with contradictory documents, timeouts, and required abstention. Track accepted completions, unsupported claims, recovery, output tokens, latency, and reviewer effort. Repeat after any provider version change.

Evidence used

Beam AI support status

Under evaluation. This page does not confirm production support, regional access, or equivalence between preview and hosted artifacts.

Use case 1

Multilingual operations agent

Run the same document-and-tool process in Chinese and English, including contradictory evidence and required escalation. Measure completion, unsupported claims, language drift, retries, latency, and reviewer time.

Use case 2

Efficient coding workflow

Compare Hy3 with a larger open-weight model on repository changes with tests and recovery from a broken tool. Track accepted patches, regressions, output tokens, wall-clock time, and accelerator cost.

Use case 3

Tencent Cloud deployment

Evaluate identity, networking, region, logs, retention, quotas, private connectivity, support, and cost around the hosted model—not just prompt quality.

Use case 4

Preview-to-production parity test

Run identical prompts and tool traces against the preview weights and intended hosted endpoint. Record divergences in outputs, safety behavior, latency, and tool-call schemas before relying on portability.

Related LLMs

Curated alternatives to compare before selecting a model family.

ChatGPT

ChatGPT is OpenAI’s conversational product; enterprise automations use the underlying GPT models and APIs. This guide separates the product from the model family and explains how to evaluate the current GPT-5.6 lineup.

ChatGPT

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

DeepSeek

DeepSeek’s current V4 family provides Pro and Flash API models with thinking and non-thinking modes, a documented one-million-token context window, and compatible API interfaces.

DeepSeek

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Perplexity

Perplexity’s Sonar API family combines language models with web search and citations. The core enterprise decision is whether provider-managed retrieval and source handling fit the workflow.

Perplexity

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Gemini

Google’s Gemini family covers production and preview models for multimodal, reasoning, audio, media, and tool-oriented workloads. Model stage and endpoint stability are central to an enterprise selection.

Gemini

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Claude

Anthropic’s Claude family spans frontier reasoning, coding, vision, and faster production models. The practical choice is a specific pinned model ID on the Claude API or a supported cloud platform.

Claude

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Llama

Meta’s Llama 4 family provides open-weight, natively multimodal models. Scout emphasizes very long context and efficient deployment; Maverick targets stronger general multimodal capability.

Llama

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Start Today

Build AI agents with the right model

See how Beam can orchestrate governed AI workflows across the model family that fits your requirements.

Start Today

Build AI agents with the right model

See how Beam can orchestrate governed AI workflows across the model family that fits your requirements.

Start Today

Build AI agents with the right model

See how Beam can orchestrate governed AI workflows across the model family that fits your requirements.

FAQs

Frequently Asked Questions

Model selection, deployment, governance, and Beam support questions answered.

What is Tencent Hy3?

Hy3 is Tencent Hunyuan’s July 2026 agent-focused MoE model with 295B total parameters, about 21B active per token, hybrid fast/slow reasoning, and a 256K context window.

How strong is Hy3 independently?

Artificial Analysis reports an Intelligence Index score of 41, about 59 output tokens per second, and roughly 2.5 seconds to first token for the tested endpoint. It is strong early evidence, not a long production history.

Is Hy3 open-weight?

Tencent has published Hy3-preview weights. Buyers should verify whether the exact production Hy3 artifact they intend to use is available, under which license, and whether it matches the hosted service. The current classification is Mixed.

What is the main Hy3 deployment risk?

Release maturity. Version changes, artifact parity, documentation, quotas, region, and enterprise controls need to be pinned and retested because the model is very new.

Does Beam support Hy3 in production?

Beam support is currently Under evaluation. No approved integration, production benchmark, or preview-to-hosted parity result is attached to this record.