The fire at the center of Hearth is qwen2.5-coder, an open model running entirely on the shop machine — no cloud, no bill, no data leaving the building. This is the record of how it was worked into something dependable: from stuck at 79% to a clean sweep, one night in the forge at a time. It learned to see, to plan its work, and to check the result before calling a job done.
2 GPUs · 14 GB VRAMqwen2.5-coder 14B100% on our own ironfired every evening
Every night, the model runs a graded battery in the forge — a fixed set of real coding jobs (write a program, find a bug, edit across files, do exact math on real data). Each one runs in a fresh sandbox and gets auto-scored against the right answer. No opinions, no vibes — either the code did the thing or it didn't. The goal is simple: the score should climb, night over night.
What the forge does
This is how a raw model becomes a tool you can actually trust with real work — not a novelty that's only right sometimes. Each nightly firing does three things:
Find the flaw
A miss marks the exact spot
A failed task pinpoints the precise mistake — not "the model is bad at code," but "it botched this one step."
Work it out
Each miss becomes a repair
We fix that one weakness — a clearer instruction, a better tool, a smarter setting — then run it again.
Prove it
The score has to climb
Same graded jobs, night over night. That's how it went from stuck at 79% to a clean sweep.
The payoff: the more it passes, the more real work you can safely put in its hands — real files, real bugs, real jobs — and the tests are how we know exactly what it's ready for, and what it isn't yet.
The tempering
On the small model79%
Where it started — and hit a wall. No amount of coaching could push it past that mark.
Moved up to the bigger model93%
The real unlock. The wall wasn't the instructions — it was the size of the model.
Fixed how it edits & re-checks work100%
A perfect run on the core set — 17 of 17.
Taught it imports & file paths19/19
The last stubborn task fell. A clean, repeatable perfect score on the full battery.
Then we made it harder26/27
Added tougher jobs — reading a crash, cross-file bugs, parsing data, building a small app from scratch. It cracked nearly all of them.
Taught it to plan & check its own workthe leap
The big one: now it writes out a plan before it touches anything, and proves the result after — the way a tradesman actually works, not just reacting step by step.
Steadied its hand27/27
Dialed the last bit of randomness out of how it works — and it cleared the entire hard set in one run. A first: 27 out of 27.
What it can do
Proven
Write & run code
Writes a program, runs it, shows the output.
Proven
Find & fix bugs
Including reading a crash and repairing it.
Proven
Edit across files
Change one thing consistently everywhere.
Proven
Exact math on real data
Reads a data file, computes the real number.
Proven
Explore a project
Searches files, maps out how it's built.
Proven
Search the web
Looks things up and reads the page.
Proven
Remembers
Keeps notes across sessions.
New
Sees images
Reads screenshots, diagrams, and photos.
New
Plans ahead
Writes out the steps before it starts.
New
Checks its own work
Proves the result before calling it done.
Proven
Recursion
Writes a function that calls itself, correctly.
Proven
Builds classes
A working object with its own state and methods.
Proven
Pattern matching
Pulls emails and patterns out of raw text.
Proven
Handles errors
Fails gracefully instead of crashing.
New this week
The fire can see now.
We gave it eyes — a second local model (also free, also on the machine). Now it can look at a screenshot, a diagram, or a photo and tell you what's in it, reading any text right off the image. As a test, we showed it a picture of its own scoreboard:
The vision model, looking at the screenshot"The app in the image appears to be a lab scoreboard... The large number shown at the top reads '18/19'... a progress indicator or score."
Straight talk
Solid
Swept the core battery clean — 27 of 27, and holds ~91-95% as it grew to 40+ task types
Recursion, classes, algorithms, regex, JSON read/write, sorting, dates, string transforms, error handling — nearly all first try
Plans before acting, then verifies its own work
Bug-fixing, incl. reading a crash
Cross-file edits, data parsing, and logic puzzles
Fast, private, and free to run
Still in the fire
Nailing the trickiest multi-step tasks every run — the last few points are subtle slips it doesn't always catch, a limit of its size, not its method
We're testing a bigger model (32B) right now to see if it closes that gap