Thinking Machines Lab releases open-weight model Inkling 2 Mini that runs on a laptop
Thinking Machines Lab has released Inkling 2 Mini, an open-weight model built on a recurrent Transformer architecture that runs its middle 12 layers three times per token, with 26.4B total parameters and 3.4B active parameters. It uses about 20GB of memory and can run locally on Macs with 32GB of memory or more.
模型发布/更新
-
Thinking Machines Lab releases open-weight Inkling 2 Mini model that runs on a laptop
Thinking Machines Lab has released Inkling 2 Mini, an open-weight model built on a recurrent Transformer architecture…
-
Tencent Hunyuan releases Hy4 open weights
Tencent Hunyuan has released the open weights of Hy4, positioned as a native multimodal flagship MoE model with 770B…
-
Google 发布 Gemini 4 Hydrogen,已上线 Gemini 应用和 Search
Google 宣布 Gemini 4 Hydrogen 正式上线,主打速度和日常任务,现已可在 Gemini 应用和 Search 中使用。该模型面向多样化的智能体工作流设计,强调快速高效执行;官方称其能更紧密地遵循指令、更准确地调用工具…
产品发布/更新
-
ChatGPT launches Space to manage pages, slides, images and files in one place
ChatGPT says Space in the sidebar is now the new home for pages, slides, images, files and uploads. With ChatGPT,…
-
browser-use 1.0 released: a single Python REPL tool replaces 24 browser actions
Browser Use releases browser-use 1.0, replacing the 24 browser actions of 0.x with a single persistent python REPL…
-
Google looks back on a decade of Duplex: how it calls businesses on users' behalf in Gemini Tasks
Google Research reviews the decade-long architectural evolution of Duplex and introduces the automatic…
论文研究
-
Anthropic mathematician Levent Alpöge announces Riemann hypothesis proof with Claude, releases 17.37 million lines of Lean code
Anthropic mathematician Levent Alpöge announced on his personal website that he proved the Riemann hypothesis in…
-
METR releases November 2026 capability measurement update: AI R&D-related capabilities double roughly every 3.6 months and are increasingly hard to measure
METR released a capability evaluation update covering Opus 5.5 and GPT-6 Astra, finding that on TH1.1 and TH2 tasks…
-
MiniMax internal model MIM achieves multiple practical cryptanalysis attacks and open-sources verification traces
The MiniMax research team used its internal research model MIM to conduct automated cryptanalysis of modern…
技巧与观点
-
Runway details how to build an agent-driven software factory
Runway described its practice of gradually handing each stage of software delivery over to agents. In October, agents…
-
DeepSeek Harness marks 100 days since release as author shares 13 practical plugin use cases
Kazike shared 13 plugin use cases on the 100th day since the release of DeepSeek Harness.
-
OpenHands details software factory using agents to auto-fix CVEs in its own images
The OpenHands team details how it hands most CVE remediation work to an agent pipeline: agents select fixes, open PRs,…