千問3.5霸榜全球開源大模型前四，10分鐘通過中級程式員5小時程式設計

#benchmarks#coding-llm#downloadsqwen-3.5qwen-3.5

💡Open-source Qwen 3.5 crushes coding benchmarks: 10min vs 5hr pro task, 1B+ downloads

⚡ 30-Second TL;DR

有什麼變化

全球開源大模型排行前四

為什麼重要

此基準測試霸主地位提升開源AI採用率，挑戰封閉模型，並以優越程式設計效率加速開發者工作流程。

下一步行動

Download Qwen 3.5 from Hugging Face and test it on your intermediate coding benchmarks.

誰應關注：Developers & AI Engineers

背景與延伸：來自公開資料，非原文內容。引用 6 個來源。

•Qwen3.5-397B-A17B is a Mixture-of-Experts model with 397 billion total parameters and only 17 billion activated parameters per token[1][2][5].
•Features a native context length of 262k tokens, enabled by Gated DeltaNet + Gated Attention hybrid mechanism[1][4].
•Achieves top GPQA Diamond score of 88.4 among open-source models and excels in embodied reasoning with 67.5 on ERQA benchmark[2][3].
•Released on February 16, 2026, under Apache 2.0 license as a native multimodal model processing text and images[4][5].

📊 競品分析▸ Show

Feature/Benchmark	Qwen 3.5	Kimi K2.5	GLM-5	Claude Opus 4.5
GPQA Diamond	88.4[3]	87.6[3]	86.0[3]	-
IFEval	92.6[3]	94.0[3]	-	-
SWE-Bench (Verified/Pro)	On par[1]	-	On par[1]	Slightly above[1]
Parameters (Total/Active)	397B/17B[5]	1T/32B[3]	-	Closed
Context Length	262k[1][4]	262k[3]	-	-

•Mixture-of-Experts (MoE) architecture: 397B total parameters, 17B active per token; 3x smaller than prior 235B-A22B but with 4x more experts plus a shared expert[1][5].
•Attention mechanism: Gated DeltaNet + Gated Attention hybrid, supporting native 262k token context (vs. 32k/131k in prior models)[1].
•Vocabulary size: 250k tokens with multi-token prediction, reducing costs by 10-60% across 201 languages[2].
•Multimodal capabilities: Native text+image processing with early fusion for video; excels in document recognition (90.8% OmniDocBench)[2][5].
•Efficiency: 19x faster decoding on 256k contexts than Qwen3-Max; 8.6x faster for standard tasks without performance loss[2].

Qwen3.5 will accelerate open-source adoption in agentic coding agents

Its SWE-Bench performance matches closed models like Claude Opus 4.5 while being efficient for local deployment at 51GB RAM[1].

Efficiency gains will lower inference costs for multimodal apps by 10-60%

Multi-token prediction and large vocabulary reduce token usage across 201 languages, combined with MoE sparsity[2].

Qwen3.5 sets new open-weight benchmark for long-context reasoning

262k native context and top GPQA Diamond score enable advanced RAG and multi-step tasks previously limited to closed models[1][3].

2026-02-16

Qwen3.5 series released, including 397B-A17B MoE model

Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.

AI 週報

閱讀本週精選 AI 大事摘要 →

同主題

探索 #benchmarks

同產品

更多關於 Qwen Chat

同來源