
PLAYER GAMES · OUR FIRST IN-HOUSE MODEL
The thread that adds Fabric.
Weft is a 32B-parameter specialist trained for one job: turn a plain-language prompt into a Fabric 1.20.1 server-side mod that compiles and boots on the first pass. On a loom, the weft is the thread each pass throws across the warp, and every pass adds fabric. The weights are open. This page is the ledger.
First-pass build + boot
22/2976%
Parameters
32B
Untuned base, same eval
0/29
Training corpus
100scompile-verified
Small, dense, and ours.
Creator Hub runs on frontier models, and they are excellent. But renting intelligence has a shape: every generation costs the same, forever, and none of the failures teach the model anything. A specialist we train ourselves inverts that: cheap to run, private to operate, and every gap it exposes becomes training data we can act on.
So we built a loom. The corpus is deliberately small and dense (hundreds of examples, not millions of scraped files), and nothing enters it unbuilt. Weft learned the real Fabric API surface from code that provably compiles against it.
BASE
An open 32B coder base
METHOD
Low-rank fine-tune, weights released openly
TARGET
Fabric 1.20.1 · server-side · Java 17
CORPUS
Hundreds of compile-verified examples
No labelers. No vibes. javac decides.
Every candidate example is generated from a plain-language spec, dropped into the real Fabric 1.20.1 Gradle template (the same toolchain Creator Hub builds against), compiled with Java 17, and booted on a real server.
Failures go back with the exact compiler diagnostics, get repaired, and are re-gated before they are stored. The oracle is deterministic, adversarial, and free. It does not get tired and it does not grade on a curve.
Gate
Compile + server boot
Signal
Exact compiler diagnostics
Repair
Errors fed back verbatim
Admission
Only code that builds
$ ./gradlew build
> Task :compileJava FAILED
error: incompatible types: RegistryKey<World> cannot be converted to RegistryKey<ServerWorld>
error: no suitable method found for teleport(...)
# 2 diagnostics captured, returned to the model verbatim
⟳ repair pass 1 of 1
$ ./gradlew build
BUILD SUCCESSFUL
$ java -jar fabric-server.jar --nogui
[Server thread/INFO]: Done! For help, type "help"
✓ example admitted to corpus
One generation. One build. One boot. No retries.
A frozen, held-out eval of 29 real mod requests, written the way a creator would write them. No eval prompt appears in the training data. Each request gets exactly one greedy generation, one Gradle build, one server boot. The mod either comes out working, or it counts as a miss.
0/29
Off-the-shelf 32B coder base
Same requests, same grounding prompt. Zero mods survived build + boot.
21/29
Previous best in-house adapter
The bar Weft had to clear on the identical frozen eval.
22/2976%
Weft-32B · this release
First-pass build + boot. Zero forbidden-mapping leaks.
For scale: the routed frontier ensemble that runs Creator Hub today scores 28/29 on this same eval, and it stays the default. Weft is not a replacement. It is the open, inspectable, improvable thread beside it.
SIMPLE
Command mods: homes, warps, kits, /fly, /ping
9/1090%
MEDIUM
Custom items and blocks, saved player data, join and death events
8/1267%
HARD
Block entities, crafting recipes, scheduled announcers, mixins
5/771%
Forbidden-mapping leaks (pre-1.20 API surfaces): zero
We don't ship hype.
The seven misses are specific API-surface gaps, not mysteries. We publish them because a failure you can name is a failure you can train against, and because a model card that only lists wins isn't a model card, it's an ad.
Gamemode shortcuts
GameMode · Treats the enum like an NBT compound.
API-SURFACE GAPWelcome event
ServerLifecycleEvents.PLAYER_DISCONNECT · Invents an event that does not exist.
API-SURFACE GAPConfig file
ServerPlayerEntity.getFoodStats() · Calls a method that does not exist.
API-SURFACE GAPScoreboard kills
PlayerCombatEvents · Invents a class that does not exist.
API-SURFACE GAPTab-list suffix
tab-list name setter · Reaches for a direct setter that does not exist.
API-SURFACE GAPRandom-teleport cooldown
coordinate / server API · Bad coordinate and server API usage.
API-SURFACE GAPMulti-homes
command suggestions · Wrong command-suggestions API shape.
API-SURFACE GAPEach miss is targetable with more compile-gated examples. That is the whole point of owning the loom.
See the ledgerDownload the weights. Run them yourself.
Weft is an open-weights specialist for server-side Fabric mods. Download a release and run it on your own hardware. Consult the model card for details on the available formats and their use.
View downloadsThe weights are yours.
Three ways to run it, all Apache-2.0: the adapter, the merged bf16 safetensors, and a GGUF quantization for llama.cpp, Ollama, and LM Studio. Each artifact was re-evaluated independently on the frozen 29-task eval: the adapter scores 22/29, the merged bf16 20/29, the GGUF 22/29. The model cards break down why, and every artifact has a published checksum. Released versions are immutable; corrections get a new version, never an overwrite.
Adapter weights
tar.gz · safetensors (LoRA) · 995.9 MB
tar.gz · safetensors (LoRA)
995.9 MB
b61f029449ad…
Apache-2.0
Merged weights
tar · safetensors (bf16, merged) · 65.54 GB
tar · safetensors (bf16, merged)
65.54 GB
c16037c68384…
Apache-2.0
Verify: shasum -a 256 -c SHA256SUMS
Every pass adds fabric.
This page is append-only, like the loom it's named for. What goes on the beam next, in the order we intend to weave it.
Close the seven misses
Targeted compile-gated examples for each named API-surface gap.
IN PROGRESSMerged full weights
A single safetensors build: no adapter step, load and go.
SHIPPEDGGUF quantization
Run Weft on your own hardware, no GPU cluster required.
SHIPPEDPULL THE THREAD
Take the weights.
Run them yourself.
Download the weights and run Weft yourself. The model card explains the available releases and how to use them.