The Mash · three stories, one thread
Model choice and task specification meet in builder workflows
New model-discovery and local-model options sit alongside evidence that task specifications matter in clinical agent evaluations.
Explore
Compare Gemini CLI’s preview model-discovery command with Ollama’s shipped support for LFM3 and d1 2B on Apple hardware. Then examine M4Bench’s control subset: an operational spec raised reward by 0.162, while skills added 0.024 on top, indistinguishable from zero.[1][2][3][4]
Create
Build a small evaluation worksheet this week. Record the models available through Gemini CLI’s preview command and the supported Ollama models separately. Add a field for the task specification before comparing agent results; M4Bench’s control subset shows why that distinction is worth tracking.[1][2][3]
Theorize
Hypothesis: as model selection becomes easier to inspect, builders may put more attention on specifying tasks before adding reusable skills. M4Bench’s control-subset result supports testing that idea, but does not establish that it holds outside those eight tasks.[1][2][3]
In the mash
Google Gemini CLI adds model discovery in preview
The v0.67.0-preview.0 release lists model discovery and an IDE companion bridge for developers using the CLI.
M4Bench evaluates clinical EHR derivation agents
M4Bench tests whether AI coding agents can reproduce clinical EHR derivations and whether clinician-reviewed skills improve their results.
Ollama adds Liquid AI models for Apple hardware
Ollama added LFM3 and d1 2B support for Apple users, while llama.cpp added the model families in pre-releases.