qwen2.5-coder
ollama pull qwen2.5-coder:<size> · sizes supported: 14b
The size listed here runs all six agent surfaces, five consecutive times each: 30 of 30 runs. The only code-specialised model on this list, and it arrived by being tried after the reasoning that excluded its whole class turned out to be wrong.
What it does, and how long it takes
Section titled “What it does, and how long it takes”Seconds are the median of the runs that passed, per surface.
| Size | Disk | Investigate | Optimize | Assess | Operate | Analyze | Plan | Median | Slowest |
|---|---|---|---|---|---|---|---|---|---|
| 14b | 9.0 GB | 29s | 34s | 21s | 17s | 10s | 1s | 20s | 37s |
Every cell is 5/5, and every cell closed on its first attempt.
Its one-second plan turn is the fastest on this list, and the shape of the whole model is in that figure beside the 29 seconds it spends investigating: asked to write a statement it answers at once, asked to go and read a database it takes longer than its general-purpose sibling of the same size.
What it needs that the defaults do not give it
Section titled “What it needs that the defaults do not give it”qwen2.5-coder:14b
Section titled “qwen2.5-coder:14b”One plan ask, which no run of its five spent. It was driven with no entry of its own, so the server offered the ask on every plan run and its plans fenced a statement without being asked. The number records the configuration the thirty runs were taken under; nothing here needed it.
Why a code model was not expected to clear this
Section titled “Why a code model was not expected to clear this”The agent surfaces are not code generation. A run reads a database through tools, holds what came back, and files a report whose claims cite the readings — the SQL it writes along the way is a means, and a model that writes beautiful SQL and never calls a tool answers nothing here. On that reasoning the coder families were set aside as a class and none was measured.
The reasoning was sound and the conclusion was wrong. This model matches qwen2.5:14b
— same generation, same size, same 9 GB — cell for cell, and beats it on assessment by two thirds
of a minute. What the surfaces ask for is instruction-following against a tool schema, and a model
tuned to emit exactly-shaped code turns out to be good at exactly-shaped tool calls.
One size is measured, so this is one data point and not a claim about coder models generally. It is recorded because it is the one that falsified the exclusion.
Method, and where these numbers stop being safe to generalise from:
methodology.md.