Saltar al contenido principal

Hy3

TencentOpen WeightApache 2.0 · Uso Comercial

Descripción

Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and a 3.8B MTP layer, developed by the Tencent Hy Team. Following the Hy3 Preview launch in late April, the team scaled up post-training with higher-quality data and RL, gathering feedback from 50+ products. Hy3 outperforms similar-size models and rivals flagship open-source models with 2-5x the parameters, with strong gains in reasoning, agentic, and long-context tasks. It uses 80 layers (plus 1 MTP layer), 64 GQA attention heads (8 KV heads, head dim 128), a 4096 hidden size, 192 experts with top-8 activated, a 256K context window, and BF16 precision. Hy3 is a hybrid-thinking model supporting configurable reasoning effort (no_think, low, high), and emphasizes production-grade tool-call and output-format stability, reduced hallucination, and reliable multi-turn intent tracking.

Fecha de lanzamiento
2026-07-06
Parámetros
295.0B
Longitud del contexto
262K
Modalidades
text

Radar de capacidades

2
general
60
coding
4
reasoning
77
scienceest.
70
agents
0
multimodal

Science se estima a partir de las puntuaciones científicas de LLM Stats o del razonamiento cuando no hay benchmarks científicos dedicados.

Rankings

Dominio#PosiciónPuntuaciónFuente
Capacidad agéntica68
48.0
LS

Puntuaciones de benchmarks (LLM Stats)

(LLM Stats (zeroeval))

Agents

WildClawBench53.6%Aut.
Toolathlon48.5%Aut.

Chemistry

SuperChem54.9%Aut.

Code

Claw-Eval68.5%Aut.
SkillsBench55.3%Aut.
NL2Repo45.6%Aut.
DeepSWE28.0%Aut.
CL-bench23.8%Aut.
CL-bench (Life)17.0%Aut.

Math

USAMO 202630.24 / 42Aut.
IMO-AnswerBench90.0%Aut.
ArXivMath52.2%Aut.
MathArena Apex38.7%Aut.
HorizonMath7.1%Aut.

Physics

PHYBench77.4%Aut.
CMT-Benchmark37.9%Aut.

Reasoning

GPQANYU + Cohere + Anthropic (2023)90.4%Aut.
BrowseCompOpenAI (2025)84.2%Aut.
MCP Atlas79.1%Aut.
SWE-Bench Verified78.0%Aut.
SWE-bench Multilingual75.8%Aut.
FrontierScience Olympiad74.8%Aut.
AA-LCR73.4%Aut.
Terminal-Bench 2.171.7%Aut.
SWE-Bench ProPrinceton NLP (2024)57.9%Aut.
Humanity's Last Exam (with tools, text-only)53.2%Aut.
Humanity's Last Exam (no tools, text-only)47.0%Aut.
APEX-Agents25.6%Aut.
FrontierScience Research21.3%Aut.

Search

DeepSearchQA91.0%Aut.
WideSearch76.4%Aut.

Índices de evaluación AA

(Artificial Analysis)

No hay datos de evaluación AA disponibles

Puntuaciones por categoría LLM Stats

(LLM Stats (zeroeval))
Math
4
Reasoning
2
General
2
Biology
90
Physics
80
Search
80
Frontend Development
80
Long Context
70
Chemistry
70
Tool Calling
70
Science
60
Agents
60
Code
60
Knowledge
50
Coding
50

Precios

Precio de entrada$0.136 / 1M tokens
Precio de salida$0.555 / 1M tokens
Precio mixto (3:1)$0.25 / 1M tokens
Precio de lectura caché$0.033 / 1M tokens

Velocidad

Tokens/seg80.8
Retraso del primer token2.83s
Tiempo hasta la respuesta27.58s

Ranking de Precios por Proveedor

Ranking de Precios por Proveedor

18 proveedores

Más barato: NanoGPTMás caro: AI/ML API
ProveedorEntradaSalida
1NanoGPTMás barato
$0.066
$0.26
2OpenRouter
$0.132
$0.528
3Kilo Gateway
$0.132
$0.528
4DevPass (LLM Gateway)
$0.132
$0.528
5Eden AI
$0.132
$0.528
6LLM Gateway
$0.132
$0.528
7TencentPRINCIPAL
$0.136
$0.555
8DeepInfra
$0.14
$0.58
9OpenCode Go
$0.14
$0.58
10Novita AI
$0.14
$0.58
11Requesty
$0.14
$0.58
12Vercel AI Gateway
$0.14
$0.58
13Jalapeno Cloud
$0.14
$0.58
14Kenari
$0.1543
$0.6286
15AIHubMix
$0.1562
$0.6248
16CrossModel
$0.16
$0.64
17OrcaRouter
$0.18
$0.59
18AI/ML API
$0.2116
$0.8464

Comparar precios entre diferentes proveedores de API para este modelo.

Fuentes externas