Tencent · Hunyuan · Reviewed 2026-09-11

Hunyuan-T1

Tencent's deep-reasoning model combining Transformer and Mamba architectures in a Mixture-of-Experts configuration for ultra-fast chain-of-thought generation.

Overall · 0–100—

Status: active
Released: 2025-03-20
Context: 28,672 tokens
Parameters: Undisclosed
Inputs: Text
Type: reasoning
Max output: 16,384 tokens
API model ID: hunyuan-t1
Reasoning: thinking
Knowledge cutoff: —
Training cutoff: —
Weights / license: Unavailable
Price / 1M tokens: $0.14 in · $0.56 out
Cached input: —

Independent benchmark scores

No snapshot score

Coding

—

No comparable score for this task lens in the selected independent snapshot.

No snapshot score

Reasoning

—

No comparable score for this task lens in the selected independent snapshot.

No snapshot score

Research

—

No comparable score for this task lens in the selected independent snapshot.

No snapshot score

Multimodal

—

No comparable score for this task lens in the selected independent snapshot.

No snapshot score

Agentic

—

No comparable score for this task lens in the selected independent snapshot.

Raw benchmark results

Source-backed rows with the original evaluation context preserved.

2 results

Benchmark	Score	Source	Evaluation notes
GPQA / GPQA DiamondChecked 2026-09-11	69.3%	Tencent Cloud	GPQA Diamond score reported in official Tencent Hunyuan-T1 release evaluations.
MMLU-ProChecked 2026-09-11	87.2%	Tencent Cloud	MMLU-Pro accuracy reported in official Tencent Hunyuan-T1 release evaluations.

Strengths

Hybrid Transformer-Mamba MoE architecture achieving 60-80 tokens/sec decoding speed
Post-training focused 96.7% compute on RL for complex logic and competition mathematics
High verified scores on MMLU-Pro (87.2) and GPQA-Diamond (69.3)

Limitations

Closed-source cloud API model
Reasoning tokens and thinking trajectory require specific parameter handling in API calls

Model evidence · official · Checked 2026-09-11

Tencent Hunyuan-T1 documentation and release notes

Official documentation for Hunyuan-T1 deep reasoning model, hybrid Transformer-Mamba MoE architecture, and capabilities.

Capability Index source

Artificial Analysis Intelligence Index v4.3.2

Underlying raw metrics and exact model configurations remain authoritative.

Capability Index source

Artificial Analysis methodology

Underlying raw metrics and exact model configurations remain authoritative.

Task lenses preserve raw scores. Radar axes use score / current leader × 100; overall is the equal-weight average of all five indices. Missing evidence is not a zero or a low rank.