01 / MONITOR
See the run as it changes.
Training and validation signals sit beside the exact model recipe that produced them.
llm.istanbul / WebGPU model studio
Build a small model. Keep the corpus close.
A guided, local-first workspace for shaping an architecture, preparing documents, training with WebGPU and returning to the same experiment later.
✓ source files stay local
✓ editable build recipe
✓ checkpoint-ready
local-first
llm.istanbul / WebGPU studioModels
local-firstengine ready
Local model workspace
Build a small model, right in the browser.
Choose a size, add your documents and open a training workspace.
Start a new model →
Import config
Local-first
WebGPU
Resumable
01Browser-nativeno remote job queue
02WebGPU runtimecompute where you work
03Portable recipeexportable configuration
04Checkpoint-readypause, return, continue
01 / Guided setup
A wizard that ends in a real workbench.
Start with decisions people can actually reason about. The studio turns them into an editable build recipe rather than hiding them behind a one-click black box.
new-model / model-sizelocal
Build configurationSet the architecture target
01 / 0417.3
MB
istanbul-notes
istanbul-notes.llmconfig.jsonTarget 17.3M
GeneratedActual size17.3M
Width384Layers6Heads6Context256Vocab8kPrecisionfp32
Source documentsPrepare the local corpus
02 / 04Drop documents or browse a folderFiles are parsed and prepared in this browser.
MD
field-notes.md184 KB · 38k tokens
readyTXT
interviews.txt2.1 MB · 442k tokens
readyJSONL
examples.jsonl760 KB · 121k tokens
readyHTML
archive.html1.4 MB · 267k tokens
readyLearning intentMake the experiment legible
03 / 04AaContinue textBalanced language modeling for prose and domain adaptation.✓
Q?Question answeringGround answers in the selected local corpus.
{ }Structured outputLearn repeated patterns, records and schemas.
↻Continue a runOpen an existing checkpoint as a new branch.
Ready to buildReview before the workbench
04 / 0417M
istanbul-notes384 width · 6 layers · 6 heads
Ready to buildDocuments4 files
Corpus868k tok
PurposeContinue text
RuntimeWebGPU
✓ Files remain local and the recipe stays editable.
02 / Local-first boundary
The browser is the machine, not a remote control.
Documents move through preparation, model training and checkpoints inside one visible boundary. The interface tells you where the work happens instead of asking for blind trust.
FSLocal source pipelineSelected files feed the tokenizer and dataset preparation path in the browser.
GPUWebGPU computeThe training workspace targets the GPU available to the current browser session.
CKReturnable experimentsCheckpoints make a model run something you can pause, reopen and branch.
browser boundary / local workspace
Build pipeline active locally
Source filesText, Markdown, JSONL, CSV and HTML
Model runtimeTokenizer, batches and WebGPU training loop
CheckpointsResume the same recipe or fork a new run
compute laneWebGPU
Corpus uploadnot part of the path
03 / Portable build recipe
The setup is an artifact, not a memory.
Architecture, training choices and intent live in one inspectable configuration. Export it, reopen it, fork it or hand it to someone without translating a pile of screenshots.
{ }ReadableThe important choices remain visible instead of disappearing behind presets.
↗ShareableA compact recipe can travel separately from the private source corpus.
⎇ForkableUse a proven configuration as the beginning of a deliberate variation.
↻ReturnableReopen the setup beside a checkpoint and understand what you were doing.
istanbul-notes.llmconfig.jsonexample build recipe
validated1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
{
"name": "istanbul-notes",
"target_params": "17.3M",
"model": {
"width": 384,
"layers": 6,
"heads": 6,
"context": 256
},
"dataset": {
"documents": 4,
"local_only": true
},
"intent": "continue_text",
"runtime": "webgpu"
} // ready to build
04 / Training workspace
After “build,” the experiment stays understandable.
Monitor the run, inspect the recipe, save a checkpoint and test generation without losing the thread that connected the corpus to the model.
02 / CHECKPOINTS
Pause without abandoning.
Save the state, return later or branch a new experiment.
03 / RECIPE
Keep the whole build legible.
Corpus, configuration and runtime remain one connected story.
04 / GENERATE
Test the model in place.
Prompt the current checkpoint without leaving the workspace.
05 / INSPECT
Read the model at a glance.
Architecture and corpus coverage remain visible beside the run.
local model workspace
Start with a size. Add the corpus. Open the workbench.
The shortest path from “I have these documents” to an experiment you can inspect, pause and understand.
Open Studio →Runs in a WebGPU-capable browser