每週 AI 簡報

🐼Pandaily•2026年2月18日較早收集於 11h

DeepSeek 測試 100 萬 token 上下文模型

Post LinkedIn

🐼閱讀原文： Pandaily

#context-window#long-context#model-testdeepseek

💡DeepSeek's 1M token context rivals top models—test for RAG breakthroughs now.

⚡ 30-Second TL;DR

有什麼變化

100 萬 token 上下文模型測試於 2 月 13 日啟動

為什麼重要

這將推動開源 LLM 在長上下文處理上的極限，實現進階 RAG 和代理應用。DeepSeek 可能挑戰 Gemini 1.5 等專有領導者，加劇競爭。

下一步行動

Test the 1M-context model on DeepSeek's web platform to benchmark long-document retrieval performance.

誰應關注：Researchers & Academics

關鍵要點

•100 萬 token 上下文模型測試於 2 月 13 日啟動
•已在 DeepSeek 網頁和 App 版本提供測試
•業界預期農曆新年發布
•目標複製去年成功

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 9 個來源。

🔑 增強重點摘要

•DeepSeek expanded its production model's context window from 128K to 1 million tokens on February 11, 2026, confirmed by user observations and community testing showing over 60% accuracy at full 1M length.[1][4][5]
•The 1M token context is available in DeepSeek's web and app versions, enabling reliable fine-grained information retrieval even for low-frequency details in ultra-long texts.[1][4]
•Testing demonstrates high effective context utilization, with accuracy remaining stable up to 200K tokens and declining gently thereafter, outperforming Gemini series models.[4]
•This upgrade is linked to DeepSeek V4 (MODEL1), featuring Engram conditional memory (confirmed) and leaked 1T-parameter MoE architecture with Dynamic Sparse Attention.[1][2]
•Industry speculation ties the rollout to a potential mid-February 2026 full V4 launch, aiming to replicate prior success with superior coding and reasoning at lower costs.[3][9]

📊 競品分析▸ Show

Model	Total Parameters	Active Parameters	Context Window	SWE-bench	API Cost (Input $/1M tokens)
DeepSeek V4	1T	32B	1M	80%+ (claimed)	$0.27[3][1]
GPT-5.2	~2T (est.)	Full	256K	78.2%	$15[3]
Claude Opus 4.5	Undisclosed	Undisclosed	200K	80.9%	$15[3]

🛠️ 技術深入

Context Window Expansion: Silently upgraded from 128K to 1M tokens on Feb 11, 2026; maintains >60% accuracy at full length with horizontal accuracy curve up to 200K tokens.[1][4][5]
Engram Conditional Memory: Confirmed O(1) hash-based static knowledge retrieval, jointly developed with Peking University.[1][2]
Dynamic Sparse Attention (DSA): Leaked mechanism with 'Lightning Indexer' reducing compute overhead by ~50% for million-token processing.[1]
MoE Architecture: ~1T total parameters, ~32B active per token (more efficient routing than V3's 37B); combines with Engram and MHC.[1][2][3]
Manifold-Constrained Hyper-Connections (mHC): Addresses training stability at 1T scale; claimed 1.8x faster inference.[1]
Other: Runs on dual RTX 4090s; open-source weights under Apache 2.0; focuses on text modeling and info compression.[3]

🔮 前景展望AI analysis grounded in cited sources

DeepSeek V4's 1M context and 1T MoE at 10-40x lower inference costs than Western models could enable economically viable long-context tasks like full codebase analysis, reducing API spend by up to 72% in hybrid workflows while challenging OpenAI/Claude dominance with open-source efficiency and coding prowess (e.g., 80%+ SWE-bench).[3]

⏳ 時間線

2026-02

DeepSeek silently expands production model context window to 1M tokens (Feb 11), begins testing in web/app (Feb 13). Confirmed via users and >60% accuracy tests.[1][5]

📎 來源 (9)

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

🐼閱讀原文： Pandaily

📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #context-window

同產品

更多關於 DeepSeek

同來源

最新來自 Pandaily

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： Pandaily ↗

這是摘要，不是原文。去看原站，或訂閱每週簡報。