The Mistral 24B that runs on a Mac mini
Mistral released a 24B-parameter model this week that genuinely fits in 16 GB of unified memory with room to spare. We tested on an M2 Pro: ~38 tokens per second on a single turn, no thermal throttling for sustained 30-minute sessions. For most people this replaces >90% of what they were paying OpenAI for.
Ollama 0.4 lands model-mixing
The new release adds first-class support for routing between models at runtime. You point Ollama at a small model for hot questions and a bigger one for the hard ones, using a routing LLM as the judge. Worth a real upgrade test if you run any local setup.
opencode gets session export
The most-requested feature since launch landed in v0.42: full session export to plain Markdown and JSON. You can now hand an opencode session to a colleague, version a session in git, or paste the transcript into a code review.
What we're watching
Three quiet pivots in the agent tooling space: a major IDE shipped an “agent pane” that targets our opencode use-case, a Linux distro vendor baked in a model router, and a public-cloud provider made its vLLM-compatible inference endpoint free for under $1/1k tokens. Worth a longer write-up next week.
Subscribe to Deep Tabs for the weekly edition.