Ollama adds Thinking Machines' Inkling 2 Mini for Apple silicon Macs
Ollama says the 26B model (5.4B active) fits in about 20GB and scores 79.2% on SWE-Bench Verified. PC support is promised, not shipped.
Open source, from Ollama releases
Most of today's stories turn on what can be verified. Benchmark scores, a Lean proof that compiles and a red-team extraction rate all rest on checks, while METR finds judgment work still lags. The infrastructure items, from stolen router keys to sandboxing and org sign-in, concern where trust in the stack is placed.
7 stories, about 9 minutes to read them all.
Ollama says the 26B model (5.4B active) fits in about 20GB and scores 79.2% on SWE-Bench Verified. PC support is promised, not shipped.
Open source, from Ollama releases
After login it applies org settings and the org's gateway or cloud provider. A built-in chart tool, easel, also arrives for some plans.
Coding tools, from Claude Code releases
In OpenAI's internal red-team test, extraction from an earlier request fell from 54.0% on GPT-6 Astra; the current-request rate is 0.9%.
Safety and security, from OpenAI
The nonprofit's updated time horizon report adds longer, open-ended tasks. Models fall short of humans on designing evaluations and scoring philosophical arguments.
Research, from MIT Technology Review
Dragos researcher says attackers stole keys from three providers, resellers logged prompts, and privacy flags such as ZDR were ignored.
Safety and security, from Ars Technica
Levent Alpöge's Lean proof compiles, mathematicians say, but humans have yet to digest it, and a separate human-only claim competes with it.
Research, from The Verge
The agent-lifecycle and sandboxing projects move to vendor-neutral governance; Google says agents resume in under 500 milliseconds.
Open source, from InfoQ