Hy3

Overview

Tencent Hunyuan Hy3 opened weights and was announced as Hy3 Preview on April 23, 2026. Tencent shipped the general-availability release on July 6, 2026, refined using real product feedback from Yuanbao, ima, and WorkBuddy, with steadier coding-agent behavior, long-document understanding, and multi-turn context handling — Tencent describes it as better suited to front-end tasks and cross-file code development. The architecture is unchanged from Preview — a 295B total / 21B active MoE with 192 experts (top-8 activation) plus a 3.8B MTP speculative-decoding layer — and the GA gains come from three months of further post-training: more post-training compute, higher-quality and more diverse task data, and finer-grained hallucination detection and constraints. It remains Apache 2.0 licensed for free commercial use.

Key capabilities

Hy3 exposes reasoning_effort with no_think, low, and high. The GA release moved to flat, non-tiered pricing on TokenHub, replacing Preview's input-length tiers; hy3-preview remains available separately with its original tiered pricing. See live pricing in the model catalog.

Architecture and efficiency

Hy3 is a dense-MoE hybrid decoder-only model: the first layer uses dense FFN, later MoE layers route each token to 8 of 192 experts, and the architecture uses sigmoid routing, QK-Norm, and GQA (64 attention heads / 8 KV heads). An MTP layer supports speculative decoding. GA did not change this architecture — the three-month gap between Preview and GA went into post-training: more compute, higher-quality and more diverse data, and finer hallucination detection and constraints.

Benchmarks

Tencent's GA disclosures focus on coding and agent gains, head-to-head comparisons with peer models, and reliability numbers from live products, rather than re-publishing STEM/long-context benchmark suites.

Coding and agent gains over Preview

Versus Preview: SkillsBench rises from 29.1 to 55.3, MathArena Apex from 12.8 to 38.7, SWE-bench Pro from 46.0 to 57.9, and NL2repo from 35.3 to 45.6. Tencent summarizes the overall agent-and-code improvement as 20%–30%.

Head-to-head comparisons

ClawEval pass³ reaches 68.5, ahead of DeepSeek V4 Pro (62.4) and Qwen3.7 Max (65.2); SkillsBench's 55.3 tops the public leaderboard; BrowseComp scores 84.2, close to GPT-5.5's 84.4. GA also reports SWE-bench Verified 78.0, SWE-bench Multilingual 75.8, and Terminal-Bench 2.1 at 71.7.

Reliability and real product deployment

Overall hallucination rate drops about 50%, and about 44% in long-document/RAG scenarios. In Tencent's own products, WorkBuddy's task success rate rises from 72% to 90% with 34% lower average latency; ima's agent system reaches 95.1% stability with nearly 19% net gain in answer quality; and the WeGame assistant's multi-turn tool-orchestration success rate climbs to 92% with hallucinations down from 4.5% to 2.8%. These are live production numbers, not offline benchmark scores.

When to use it

• Agent workflows: office and life automation (data processing, document generation, research analysis, information lookup, web page creation), with a disclosed 90% task success rate.
• Coding assistants: cross-file code reading/editing, debugging, and front-end tasks, with clear gains over Preview on SWE/Terminal-style benchmarks.
• Long-document processing: 256K context and 192K max input for contracts, research reports, and long code contexts.
• Reliability-sensitive use cases: sharply lower hallucination rates suit customer support and knowledge-base Q&A.
• Chinese product integrations: Yuanbao, ima, WorkBuddy, QQ, and Tencent Docs already validate low-latency, high-success-rate deployment.

CrossModel exposes Hy3 GA through OpenAI-compatible /v1/chat/completions and Anthropic-compatible /v1/messages; the earlier hy3-preview remains available as a separate configuration. Current pricing is available in the model catalog.