Every model, test by test.
The whole lineup, run through the same tests on the same laptop processor with no graphics card. Pick a model to see its results. The numbers are generated from the raw results kept with the app's source, so this page can't round anything up.
Not comparable with published scores. Each test is a small, repeatable sample of the real benchmark, asked once at temperature 0. Large labs test on the full sets, with their own prompts and often many attempts. Nothing here is a claim against any model that wasn't run, and on samples this small two models a few points apart are a tie: read the ranges, not the decimals.
All models side by side
Share of each sample answered correctly. Sort by any test; pick a row to open that model above. A dash means the test hasn't been run on that model yet, and the line under the table says why.
How it was measured
Six standard tests, on a sample. MMLU (40 questions), ARC-Challenge (30), HellaSwag (25), TruthfulQA (25), GSM8K (20) and HumanEval (20), from the public datasets, asked at temperature 0 with thinking off. A processor writes about 15 tokens a second, so the full sets would take weeks per model. The interval beside each score is the range the full-set score probably falls in (95%, Wilson). HumanEval runs the Python each model wrote, in a temporary folder, with a 10-second limit per task.
Describe an app, get one that works. Twelve everyday requests go to the Code tab's app maker with the model under test. Each result is opened in a real browser and checked the way a person would: it loads without an error, and its controls do what was asked. "After one fix" is what the Code tab does when a running app reports an error: the error goes back to the model once.
Where they ran. An Intel Core Ultra 7 258V laptop with 31 GB of memory and no graphics card. Power models run through Ollama; Trixe models through LCOA's own llama.cpp engine, from the pinned files the app downloads, or for the small ones from the official files listed in the repository. The app's token counter is checked against each model's own count on every run.
Response times
Measured separately on an older desktop: an Intel Core i3-7100 with two cores and 16 GB of memory, processor only. The probes ask a short question and allow up to eight words of answer, so they test how fast a model starts talking, not how good its answers are.
First words of a new question, model already loaded
Cold start and warm, in seconds
Models that keep a snapshot of their instructions
Five models can't reuse a cached prompt the usual way, so LCOA keeps a snapshot of their state in memory, capped at 512 MB and released when memory runs low. Total time for the same question, fresh against restored.
Tool use
On the same older desktop, each small model was offered three real read-only tools (a computer report, the task list in an empty test profile, and the list of open windows) and asked for each in turn. A pass means it chose the right tool and the call ran.
Separately, all ten wrote a valid read-only PowerShell call and ran it through the guarded tool, 10 of 10. A compound command that would have written something was rejected, as it should be.
Nano chose the computer report for every question, one reason Triangle never picks it automatically. T1 ran past the three-minute limit for a structured call. Passing three checks doesn't mean a model gets every task right.
Processor or graphics?
The test laptop also has an Intel Arc 140V graphics chip. Run through Vulkan, it reads a prompt seven times faster than the processor but writes nine times slower, and writing is most of the wait. So LCOA keeps models on the processor. Thread counts and cache settings made no difference worth having.
Qwen3 4B Instruct reading a 1,463-token prompt and writing about 170 tokens, 23 October 2026. The graphics test ran in a second copy of Ollama, stopped afterwards. Intel's own builds of the engine (IPEX-LLM) may do better on this chip; they weren't tried.
Not measured
Leaving these out is a choice, not an oversight. A half-run of a hard benchmark would mean nothing, and a number on this page should mean something.
Benchmarks that need long reasoning, tools or a hosted grader
MATH, GPQA, MMLU-Pro, BBH, IFEval, LiveCodeBench, SWE-bench, and agent and computer-use suites. A model this size on a processor can't finish them in reasonable time.
Models that couldn't run here
Worm helpers
All ten load and generate: twenty runs, each specialist in both of its modes. That shows they work, not that they're right, and in testing several weren't. They're used only when you ask for one, and checked by the main model.
Comparisons with other models
No scores are given for any model that wasn't run here, and none of these numbers should be set beside a published leaderboard. Different questions, prompts and attempts make that comparison meaningless.
Run them yourself
One command downloads the lineup (each file checked against its published hash) and runs every test on every model, keeping what is already done. Every answer each model gave is kept alongside its score.
node scripts/bench-lineup.mjs node scripts/bench.mjs --model "Trixe Pro" node scripts/vibe-bench.mjs --model "Trixe Pro"