AI Tracker下載 App

模型新聞，
為開發者整理。

各家 AI 公司最新的模型動態：發表、調價和停用，並附上對你的影響。

傳聞Moonshot AI· 模型· 8 則討論

Moonshot 以影片預告 Kimi K3.1；尚未正式發布該模型

Moonshot 的 Kimi 帳號發布了一段 24 秒影片，未附說明文字。引用該影片的帳號描述，影片中 Kimi Code 終端顯示「Model: Kimi K3.1」，並用 Three.js 編寫程序化 3D 場景。Moonshot 尚未公布該模型、發布日期、定價或權重。MiniMax· 模型· 7 則討論

MiniMax 公開 MiniMax M3.1 Flash 權重

MiniMax 表示，MiniMax M3.1 Flash 的權重現已在 Hugging Face 和 ModelScope 上公開。該模型於 10 月 19 日推出時，MiniMax 曾表示權重將於日後釋出。模型· 1 則討論

Artificial Analysis 發布 Coding Agent Index v2.0；Claude Code 搭配 Opus 5.5 以 61.7 領先

Artificial Analysis 推出 Coding Agent Index v2.0，這項指標會將模型與其 agent harness 一併評分，涵蓋軟體工程、終端機使用和程式碼庫問答等 290 項任務，權重均等。分數由 Artificial Analysis 自行評定，未經獨立驗證。模型

Perplexity 推出 Mac 版 Decider，可在本機執行其 4B 決策模型

Perplexity 表示，現在你可以在 Mac 上執行 pplx-decider-flash-v1.2，無須 API 金鑰。Decider for Mac 是一款可在裝置上執行 4B 決策模型的 App：你可以針對文字或圖片提出問題，並在自己的機器上取得機率值。OpenAI· 模型· 3 則討論

OpenAI 的 Decisions API 現已正式推出

OpenAI 表示，於 10 月 6 日進入公開測試版的 Decisions API 現已正式推出。據 OpenAI 表示，應用程式和代理程式可用它來分類文字與影像、評估輸入內容，並決定接下來要做什麼，而無須生成完整的模型回應。在測試版公告中，OpenAI 表示，透過 Responses API，Decisions API 的決策速度最高可達 GPT-6 Luna 的 10 倍。Anthropic· 模型· 2 則討論

Anthropic 表示，Claude 為 123 個開源函式庫撰寫了形式化證明，證明它們不會崩潰

Anthropic 表示，該公司與 Theorem 合作，使用 Claude 在組合語言層級（RISC-V）證明 100 多個真實世界開源函式庫的安全性質；Anthropic 稱這是 AI 首次大規模用於在該層級形式化驗證正式環境軟體。OpenRouter· 模型

OpenRouter 的目錄為 Arcee AI: Trinity Large Thinking 標示了 10 月 26 日的到期日

OpenRouter 的模型目錄現在為「Arcee AI: Trinity Large Thinking」（arcee-ai/trinity-large-thinking）標示到期日為 2026 年 10 月 26 日。此模型自 4 月 1 日起列入目錄，脈絡長度為 262,144 個 token。該條目先前沒有到期日。這是目錄欄位；據我們所見，OpenRouter 和 Arcee 都未說明原因，條目也未說明模型在該日期之後會如何處理。模型

Ideogram introduces its first vector model

Ideogram announced its first vector model on X. According to Ideogram, it generates clean, editable SVGs with smooth gradients and sharp typography from a text prompt, or turns any image into a detailed, scalable vector.模型

Ling announces Ling-3.1-flash-1019, with API access and open weights to follow

The Ling team announced Ling-3.1-flash-1019 on X, an update to its Ling-3.1-flash model. The figures below are the team's own claims, based on Artificial Analysis benchmarks.Mistral AI· 模型

Mistral releases Voxtral 3, speech models and a voice agent SDK

Mistral announced Voxtral 3 on October 20, 2026 (shared on Hacker News, where it had 735 points and 182 comments by October 20). Details and benchmarks below are Mistral's own claims from its announcement page.DeepSeek· 模型

Vercel adds DeepSeek V4.1 Flash 1017 to Token Pass

Vercel says DeepSeek V4.1 Flash 1017 is now included in Token Pass, its subscription for AI model tokens. Vercel's post lists the new weights from DeepSeek, a 1M-token context window and no configuration change, and advertises $80 of tokens for $30 a month.Anthropic· 模型· 5 則討論

Anthropic opens a research preview of Claude Duet, a proactive agent app

Anthropic has announced a research preview of Claude Duet, a new app where Claude works as a proactive agent. According to Anthropic, it brings you what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.OpenCode Go· 模型

OpenCode adds Meta's Muse Wonder model

OpenCode announced on X on October 21 that Meta's Muse Wonder model is now available in OpenCode. The post gives no further detail, such as which variant, pricing or how to select it. A minute earlier the account had posted the same announcement under the name "Meta Wonder".模型· 1 則討論

Recursive releases Flint, an open-weight local coding model

Recursive (@Recursive_SI) announced Flint on X on October 21, calling it its first model: a fast coding model meant to run on a user's own machine. The post, by Richard Socher, says it was built with the company's automated research system, the Eureka Machine, and from scratch in under four months.Z.ai· 模型

Z.ai's pricing page adds GLM-5.3-FlashX-Legal

Z.ai's API pricing page now lists a model named GLM-5.3-FlashX-Legal. The page gives its price as $0.37 per 1M input tokens and $1.25 per 1M output tokens, with cached input at $0.075 per 1M tokens.提醒OpenAI· 模型· 5 則討論

OpenAI says its models helped find critical security flaws in PHP

OpenAI announced that its models helped researchers, working under its "Patch the Planet" initiative, discover critical security flaws in PHP, which it says runs an estimated 70% of websites. According to OpenAI, it led a coalition of maintainers, researchers and security providers for about two months to prepare fixes and protections before disclosure; the fix was intricate and required a major change to the PHP engine.OpenAI· 模型

OpenAI describes using thousands of AI agents to find new leads in biological data

OpenAI published a post titled "Using thousands of AI agents to uncover new leads in biological data" (page address: "scaling-ai-research-for-complex-biological-data"). By its title, OpenAI says it ran thousands of AI agents on complex biological datasets to surface new research leads. That is OpenAI's own account.OpenAI· 模型· 4 則討論

OpenAI says its models are coming to Google Cloud

OpenAI has published a post titled "OpenAI models are coming to Google Cloud". Its summary says OpenAI and Google Cloud are bringing frontier models and Codex workflows to Gemini's Model Garden, available through Google Cloud Marketplace. This is OpenAI's own account. Posts on X about the announcement say the models and Codex arrive on Google Cloud in the coming weeks.DeepSeek· 模型

Stork: Alibaba raised its price for DeepSeek V4 Flash 0731 by 230%

Stork AI Daily, a newsletter that runs its own LLM price catalog, reports that Alibaba raised the price of DeepSeek-V4-Flash-0731 overnight. This is Stork's account from its own tracker; neither Alibaba nor DeepSeek has announced it, and the newsletter does not say which Alibaba service it means.模型· 1 則討論

KRAFTON AI releases Raon-1 mini preview, an open 30B mixture-of-experts model

KRAFTON AI announced a preview of Raon-1 mini on October 20, a mixture-of-experts language model trained from scratch with about 30B parameters in total and 4B active. The weights are on Hugging Face as KRAFTON/Raon-1-mini-preview under Apache 2.0.
較早的新聞