Hunyuan T1

Summary

Hunyuan T1 is Tencent's flagship reasoning model, officially released mid-February 2026 — billed as the world's first ultra-large-scale Hybrid-Mamba-Transformer MoE. It scores 87.2 on MMLU-PRO and 96.2 on MATH-500 and is API-only via Tencent Cloud.

Overview

Hunyuan T1 is Tencent's flagship reasoning model, with the official version released in mid-February 2026 following an earlier T1-Preview phase. Architecturally, Hunyuan T1 is one of the most distinctive models in the 2026 frontier landscape: it is built on the TurboS base, marketed by Tencent as "the world's first ultra-large-scale Hybrid-Mamba-Transformer MoE" reasoning model. The Mamba state-space-model component delivers efficient long-context handling at lower per-token compute than pure transformer architectures, making T1 a meaningful architectural departure from the otherwise transformer-dominant frontier model landscape.

T1's benchmark profile is competitive with the global frontier: 87.2 on MMLU-PRO (second only to OpenAI's o1 at the time), 96.2 on MATH-500 (closely trailing DeepSeek R1), 64.9 on LiveCodeBench, and 91.9 on ArenaHard. The model is API-only as of April 23, 2026, distributed through Tencent Cloud and select integration partners — distinguishing it from open-weight Chinese competitors like DeepSeek V4, Qwen 3.5, and Moonshot Kimi K2.6.

Specifications

  • Developer: Tencent (Hunyuan AI division)
  • Release Date: T1-Preview earlier in 2026; official T1 mid-February 2026
  • Type: Frontier reasoning large language model
  • Architecture: Hybrid-Transformer-Mamba MoE (TurboS base) — claimed first ultra-large-scale Hybrid-Mamba-Transformer MoE
  • Distribution: API-only (Tencent Cloud); not open-weight
  • Strategic Positioning: Frontier reasoning at architectural distinction from pure-transformer competitors

Capabilities

Hybrid-Mamba-Transformer Architecture (defining capability): Combines transformer attention with Mamba state-space-model components. Mamba components reportedly deliver efficient long-context handling at lower per-token compute than pure-transformer approaches — a meaningful architectural alternative to the GPT/Claude/Gemini playbook.

Strong Reasoning Benchmarks:

  • 87.2 on MMLU-PRO (second only to OpenAI's o1 at release)
  • 96.2 on MATH-500 (close to DeepSeek R1)
  • 64.9 on LiveCodeBench
  • 91.9 on ArenaHard

Project-Level Code Generation: Subsequent updates added enhanced project-level code generation skills.

Improved Text Generation Quality: Updates have also focused on text generation and writing quality improvements.

Better Multi-Turn Comprehension: Recent updates address multi-turn text comprehension — a known weakness in many reasoning models that overemphasize single-turn benchmark scores.

API Distribution via Tencent Cloud: Integrated with Tencent's broader cloud platform; available for enterprise customers through Tencent's existing cloud relationships.

Limitations

API-Only (Closed): Unlike Tencent's earlier Hunyuan releases and unlike major Chinese open-weight competitors (DeepSeek, Qwen, Moonshot), T1 is not open-weight. Self-hosting is not available.

Geopolitical Considerations: As a Tencent product, Hunyuan T1 deployment in U.S. enterprise environments often involves additional review around data handling, compliance, and ByteDance/TikTok-style scrutiny applied to other major Chinese tech companies.

Behind Absolute Frontier: T1's MMLU-PRO 87.2 and MATH-500 96.2 are top-tier but trail the absolute frontier represented by GPT-5.5 (April 2026), Claude Opus 4.7 (April 2026), and Gemini 3.1 Ultra. The architectural distinctiveness is more interesting than absolute benchmark leadership.

Mamba Architecture Maturity: While the Hybrid-Mamba-Transformer architecture is a meaningful research direction, the broader ecosystem (training tools, inference optimizations, downstream fine-tunes) is less mature than for pure-transformer models — limiting third-party innovation around T1.

Recent Developments

  • Mid-February 2026 Official Release: Upgraded from T1-Preview to the official version with significantly enhanced reasoning capabilities.
  • Subsequent Updates: Added project-level code generation, improved text generation quality, and better multi-turn comprehension.
  • API-Only Posture (as of April 23, 2026): T1 distributed exclusively through Tencent Cloud — a notable closed posture distinct from earlier Hunyuan releases.
  • TurboS Architecture: Introduced earlier in March 2026 as the world's first ultra-large-scale Hybrid-Mamba-Transformer MoE — the architectural foundation for T1.
  • Industry Position: Strong reasoning benchmarks place T1 in the top tier of Chinese frontier models alongside DeepSeek R1 / V4-Pro, Qwen 3.5, and Moonshot Kimi K2.6, with architectural distinctiveness as the differentiating factor.

Last Updated

May 7, 2026

→ Back to Models