From the meLink team
Category: Open Source AI
-
Fable 5.1 Solved a Cipher. Now It Needs a Test
Fable 5.1 reportedly solved a centuries-old cipher. The real lesson for teams is how to make agent investigations reproducible and…
-
Real-SWE Benchmark Puts Coding Agents on Notice
Real-SWE tests coding agents on private production codebases. Its early scores show why verification matters before teams delegate real work.
-
Anthropic Threat Report: An Agent Ops Warning
Anthropic’s new threat report shows why capable agents need permission boundaries, audit trails and real human checkpoints.
-
DeepSeek V4.1 Flash Shrinks Agent Memory Costs
DeepSeek V4.1 Flash targets a costly agent problem: retaining long context without making memory infrastructure the bottleneck.
-
Desert Ant Local Models Move AI Onto Everyday Devices
Desert Ant local models put on-device AI back in focus, giving builders a concrete reason to separate private, fast tasks…
-
MiniCPM5-2B Brings Tool Use to Local Agents
MiniCPM5-2B brings long context and tool use to local AI agents. Here is what builders should test before deploying it.
-
DeepSeek V4 Flash Vision Brings Eyes to Agents
DeepSeek V4 Flash Vision brings image understanding into agent workflows. The real opportunity is safer visual evidence, routing and handoffs.
-
Git-Native Agent Memory Gets an MCP Server
OKF Agent Memory puts persistent agent knowledge in Git files, bringing MCP access, review history and clearer ownership to agent…
-
Claude Formalizes Fermat’s Last Theorem in Lean
Claude’s 11-day Lean formalization of Fermat’s Last Theorem shows why verifiable agent work may matter beyond mathematics.
-
GPT-6 Astra Turns Agent Memory Into a Map
GPT-6 Astra’s ARC-AGI-3 results and K2 Horizon’s open model fleet point to a sharper question: how do agents keep context…