每週 AI 簡報

🔥36氪•2026年2月25日較早收集於 30m

阿里發布三款中型千問3.5新模型

Post LinkedIn

🔥閱讀原文： 36氪

#open-source#model-benchmarks#low-cost-inferenceqwen3.5alibabaqwen3.5qwen3.5-35b-a3bgpt-5-mini

💡Mid-size open LLMs beat GPT-4o mini, inference at $0.03/MTok equiv

⚡ 30-Second TL;DR

有什麼變化

開源Qwen3.5-35B-A3B、Qwen3.5-122B-A10B、Qwen3.5-27B

為什麼重要

提供高性能開源模型低成本，讓開發者能匹敵專有LLM而無高推理費用。提升全球建構者的AI可及性。

下一步行動

Test Qwen3.5-Flash on AliCloud百炼 for 0.2 yuan/MTok inference today.

誰應關注：Developers & AI Engineers

關鍵要點

•開源Qwen3.5-35B-A3B、Qwen3.5-122B-A10B、Qwen3.5-27B
•多榜單超越Qwen3-235B-A22B及GPT-5 mini
•Qwen3.5-Flash在阿里雲托管，每百萬Token輸入0.2元
•繼Qwen3.5-397B-A17B開源之後

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 5 個來源。

🔑 增強重點摘要

•Qwen3.5 series introduces native multimodal capabilities, unifying text, vision, and UI interaction in a single model trained on images, UI screenshots, and structured content for tasks like visual question answering and pixel-level grounding[1][2].
•Qwen3.5-397B-A17B, recently open-sourced prior to mid-size models, ranks #3 on Artificial Analysis Intelligence Index with 45 score, behind GLM-5 (50) and Kimi K2.5 (47), featuring 262K context window and 17B active MoE parameters[2].
•Qwen3.5-Plus hosted variant offers 1M token context window, 'Auto' mode with adaptive thinking and tools like search and code interpreter, outperforming Qwen3-Max in multimodal tasks at lower cost[1][3][4].

📊 競品分析▸ Show

Feature/Model	Qwen3.5-397B-A17B	GLM-5	Kimi K2.5
Intelligence Index Score	45 (#3 open weights)	50	47
Active Parameters (MoE)	17B	32B	32B
Output Tokens (Index)	~86M	110M	89M
Native Vision	Yes (image/video)	Not specified	Not specified

🛠️ 技術深入

•Qwen3.5-397B-A17B: 397B total parameters, 17B active (MoE architecture), 262K token context window, supports reasoning and non-reasoning modes in one model, native image/video input[2].
•Qwen3.5-Plus: 1M token context window (extended), multimodal (text/image/video), includes 'Thinking', 'Fast', and 'Auto' modes with built-in tools (search, code interpreter), Max CoT up to 81,920 tokens[1][3][4].
•Qwen3.5-397B-A17B hardware: FP16/BF16 requires ~800GB VRAM; 4-bit quantized ~220GB unified memory (runnable on Mac Studio/Pro M-series Ultra or multi-GPU setups like 3x A100 80GB)[1].
•Unified training on text, images, UI screenshots, structured content; 19x faster decoding on 256K long-context vs. Qwen3-Max, 8.6x faster standard workflows without intelligence loss[1].

🔮 前景展望AI analysis grounded in cited sources

Qwen3.5 mid-size models will accelerate open-source MoE adoption in mid-tier deployments

Their outperformance of larger priors like Qwen3-235B-A22B at lower active parameter counts (e.g., 3B/10B) enables efficient scaling for resource-constrained users[1][2].

Alibaba Cloud pricing at 0.2 yuan/million tokens will capture 20% more enterprise API market share

Low-cost Flash hosting combined with 1M context and tools positions it competitively against pricier multimodal APIs from Western providers[1][3].

Native multimodal unification in Qwen3.5 will become standard for Chinese open models by mid-2026

First to merge text/vision lines ahead of separate Qwen3-VL, following industry trend and outperforming prior VL series[1][2].

⏳ 時間線

2025-09

Qwen3 series released, establishing baseline with separate text and vision models

2025-09-11

Qwen3 initial release on GitHub

2025-12-01

Qwen-Plus (Qwen3 series) snapshot released with 32K context and batch pricing

2026-01-23

Qwen3-Max snapshot with integrated thinking modes and tool support (search, code interpreter)

2026-02-15

Qwen3.5-Plus snapshot released with 65K context, thinking by default, and multimodal enhancements

2026-02-16

Qwen3.5 series launched on GitHub with first 397B-A17B MoE model open-sourced

📎 來源 (5)

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

🔥閱讀原文： 36氪

📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #open-source

同產品

更多關於 Qwen Chat

同來源