Inquiries · open · since 6 Oct 2026 · last tended 16 Oct 2026
What can a local model do on a personal machine?
On a personal machine, what limits a local model is less the model than the method around it: examples, a stable prompt, and a benchmark replayed on the same answers.i
- det. 16 Oct · correctedCorrected: the launcher was reframed on 6 October but renamed LLMarium only on 10 October; listing each model with the process that launched it dates from 9 October, not 13 October.
- det. 16 Oct · interpretedWorking formulation: the method around the model matters more than the model. To be tested, not yet shown: v3 changed method and benchmark at once.
- det. 15 Oct · measuredMethod v3 on the Mac: 57 %, on a changed and larger benchmark (186 choices). Not comparable with the runs before.
- det. 14 Oct · measuredThe same task on the Mac with a local model: 32 % with 30 examples, 38 % with 69, in 5 to 8 minutes.
- det. 13 Oct · measuredCloud agents, with Julien’s profile and 69 examples: 45 % of reasons exact on 138 choices, against a floor of 23 %.
- det. 6 Oct · interpretedStarting point: a launcher for local models, reframed and renamed LLMarium. Which models are in memory, and what launched them?
Position
Measure before choosing. A local model is judged on the user’s own answers, replayed at every change of model or method, with the floor and the ceiling shown next to the score.i
What it refuses to claim
- A local score is not a ranking of models: it holds for one task, one person’s answers and one machine.
- Faster or cheaper is not better until the same benchmark says so.
- Privacy and cost are reasons to run locally, not results.
Working concepts
- Replayed benchmark
- the same question, with the same truth, asked again whenever the model or the method changes; a change of protocol is shown as a break.
- Floor and ceiling
- what a trivial answer scores (23 %), and how often two independent judges agree (75 %): a score means something only between the two.
Experiments
- done
Apple Silicon for tabular models. Make NVIDIA’s structured-data-models run on the Mac (mps and mps:0 were two devices). Merged upstream.
- running
The intentions benchmark. Why was a link kept? Five runs from 13 to 15 October, cloud then local.
- proposed
Method or model. Replay method v3 with the cloud agents: does 57 % come from the method or from the local model?
- proposed
A stable prompt prefix. Do agent tools (Codex, Pi, Zcode, OpenCode) keep the start of the prompt stable, and what does instability cost locally? Cache reuse measured per tool on one task.
What it could lead to
- already feeding LLMarium: which local models are in memory, and what launched them. · LLMariumi
- already feeding Spicilège: a local model proposes why each item was kept and which notebook it belongs to. · Spicilègei
Reading
Seven finds in the Local AI workshop notebook.
Kept by Julien; the reasons were proposed by a model on 13 October, not confirmed, to re-check on 13 November.
Crossings
In the register
- correction LLMarium
The local model launcher is reframed and renamed LLMarium (6 October).The launcher was reframed on 6 October and renamed LLMarium on 10 October. c - correction LLMarium
Models in memory are listed with the process that launched them (13 October).Listing each model in memory with the process that launched it was committed on 9 October. c - question Research notes
- measure Spicilège
- det. 15 Oct · measured57 % with method v3, which merges two verbs, on 186 choices.
- det. 14 Oct · measured38 % with a local model on the Mac, same examples.
- det. 13 Oct · measured45 % with Julien’s profile and 69 of his choices as examples.
- det. 13 Oct · measured26 % of reasons guessed right from the content alone, on 138 of Julien’s choices.
- milestone LLMarium
Models in memory are listed with the process that launched them. m
- milestone Spicilège
A local model proposes bookmarks for notebooks; tags that evolve. m
- decision LLMarium
The local model launcher is reframed and renamed LLMarium. m
- contribution NVIDIA SDM