Passer au contenu principal

Qwen3.6 27B (Reasoning)

AlibabaQwenOpen WeightApache 2.0 · Usage Commercial

Description

Qwen3.6-27B is a dense 27-billion-parameter multimodal model in the Qwen3.6 series, supporting both vision-language thinking and non-thinking modes in a single unified checkpoint. The 64-layer language model uses a hybrid layout of 16 repeats of (3 × Gated DeltaNet → FFN, 1 × Gated Attention → FFN) with hidden dim 5120 and FFN intermediate 17408 — Gated DeltaNet has 48/16 heads for V/QK (head dim 128) and Gated Attention has 24/4 heads for Q/KV (head dim 256). It supports a native 262,144-token context extensible to ~1,010,000 via YaRN and is trained with multi-token prediction. The release delivers flagship-level agentic coding, surpassing the previous-generation open-source flagship Qwen3.5-397B-A17B (397B total / 17B active) on every major coding benchmark including SWE-bench Verified (77.2), SWE-bench Pro (53.5), Terminal-Bench 2.0 (59.3), and SkillsBench (48.2), and reaches 87.8 on GPQA Diamond. Released as open weights under Apache 2.0; accessible via Qwen Studio with the Alibaba Cloud Model Studio API coming soon.

Date de sortie
2026-04-22
Paramètres
27.8B
Longueur du contexte
262K
Modalités
image, text, video

Radar de capacités

80
general
60
coding
80
reasoning
77
scienceest.
60
agents
80
multimodal

Science est estimé à partir des scores scientifiques de LLM Stats ou du raisonnement lorsque les benchmarks scientifiques dédiés ne sont pas disponibles.

Classements

Domaine#RangScoreSource
Capacité agentique135
35.0
LS
Classement multimodal16
65.0
LS

Scores de benchmarks (LLM Stats)

(LLM Stats (zeroeval))

Agents

AndroidWorld70.3%Aut.

Code

QwenWebBench1487.00 / 2000Aut.
Claw-Eval60.6%Aut.
ZClawBench53.4%Aut.
SkillsBench48.2%Aut.
NL2Repo36.2%Aut.

General

C-Eval91.4%Aut.

Language

MMLU-Redux93.5%Aut.
MMLU-Pro86.2%Aut.

Math

AIME 202694.1%Aut.
HMMT 202593.8%Aut.
HMMT2590.7%Aut.
MathVista-Mini87.4%Aut.
DynaMath85.6%Aut.
HMMT Feb 2684.3%Aut.
IMO-AnswerBench80.8%Aut.

Multimodal

VideoMMMU84.4%Aut.
MMMU82.9%Aut.

Reasoning

GPQANYU + Cohere + Anthropic (2023)87.8%Aut.
LiveCodeBench v683.9%Aut.
CharXiv-R78.4%Aut.
SWE-Bench Verified77.2%Aut.
SWE-bench Multilingual71.3%Aut.
SuperGPQA66.0%Aut.
Terminal-Bench 2.0Stanford × Laude Institute (2026)59.3%Aut.
SWE-Bench ProPrinceton NLP (2024)53.5%Aut.
Humanity's Last Exam24.0%Aut.

Video

MLVU86.6%Aut.

Vision

CountBench0.98 / 100Aut.
VLMsAreBlind97.0%Aut.
V*94.7%Aut.
RefCOCO-avg0.93 / 100Aut.
MMBench-V1.192.3%Aut.
OCRBench89.4%Aut.
VideoMME w sub.87.7%Aut.
EmbSpatialBench0.85 / 100Aut.
RealWorldQA84.1%Aut.
MMStar81.4%Aut.
CC-OCR81.2%Aut.
MMMU-Pro75.8%Aut.
MVBench75.5%Aut.
RefSpatialBench0.70 / 100Aut.
ERQA62.5%Aut.
SimpleVQA0.56 / 100Aut.

Indices d'évaluation AA

(Artificial Analysis)

Aucune donnée d'évaluation AA disponible

Scores par catégorie LLM Stats

(LLM Stats (zeroeval))
Language
90
Long Context
90
Biology
90
Legal
80
Math
80
Multimodal
80
Physics
80
Reasoning
80
Spatial Reasoning
80
Structured Output
80
Embodied
80
Finance
80
Frontend Development
80
General
80
Grounding
80
Healthcare
80
Chemistry
80
Text-to-image
80
Video
80
Vision
80
Image To Text
70
Economics
70
Code
60
Tool Calling
60
Agents
50

Tarification

Prix d'entrée$0.6 / 1M tokens
Prix de sortie$3.6 / 1M tokens
Prix mixte (3:1)$1.04 / 1M tokens

Vitesse

Tokens/sec55.9
Délai du premier token3.52s
Temps de réponse104.98s

Classement des Prix par Fournisseur

Classement des Prix par Fournisseur

8 fournisseurs

Moins cher: DevPass (LLM Gateway)Plus cher: Novita
FournisseurEntréeSortie
1DevPass (LLM Gateway)Moins cher
$0.12
$1.08
2DeepInfra
$0.32
$3.2
3Venice AI
$0.325
$3.25
4EmpirioLabs AI
$0.41256
$2.47538
5Cortecs
$0.446
$3.008
6OVHcloud AI Endpoints
$0.47
$3.19
7AlibabaPRINCIPAL
$0.6
$3.6
8Novita
$0.6
$3.6

Comparer les prix entre différents fournisseurs API pour ce modèle.

Sources externes