You can run a 26B model scoring 79.2% on SWE-Bench Verified locally on a Mac at about 20GB, and Ollama on Apple silicon now serves concurrent requests, which matters for running multiple local agents.
Key takeaways · AI-distilled
Ollama v0.40.17 adds Thinking Machines' Inkling 2 Mini: 26B parameters with 5.4B active, looped layers for more compute per tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → than its size suggests, about 20GB on a Mac, and 79.2% on SWE-benchThe standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.Full definition → per the notes. This release supports it on Apple silicon only.
`ollama launch claude` now supports Claude Code's deferred tool loading, so MCPThe Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.Full definition → tools are loaded when they are needed instead of all at once.
On Apple silicon, Ollama now serves multiple requests at once. Separately, disabling thinking now maps to the model's lowest supported reasoning level instead of being ignored.
Tool definitions sent through the API now keep JSON schema keywords Ollama doesn't model instead of dropping them, and Vulkan now uses mmap on integrated GPUs to fix 'Failed to allocate pinned memory' load failures.
Terms in this piece · Glossary
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.