Supermix is the active monorepo for the Supermix desktop app, the Omni Collective training line, specialist model experiments, benchmark tooling, and local-first packaging.
It is intentionally a mixed workspace: source code, training scripts, build helpers, generated manifests, research outputs, and packaging metadata all live in the same repository so the full desktop + training workflow can be developed, tested, packaged, and published from one place.
Supermix Studio X — V48 Frontier Edition is the latest desktop release.
The large installer and model bundles are hosted on Hugging Face due to repository asset size constraints and Git LFS quotas. The desktop release page links to the installer, model directory, and release assets.
-
source/
Active development workspace: model definitions, training scripts, dataset builders, benchmark runners, desktop/web app code, and release helpers. -
runtime_python/
A lighter packaged runtime path for quick local use. -
datasets/
Local conversation, coding, reasoning, science, and specialist training inputs. -
output/
Generated graphs, summaries, manifests, previews, logs, and Hugging Face upload staging. -
installer/
Inno Setup definitions and post-install notes for the desktop app. -
dist/
Locally built EXEs and installer outputs. -
web_static/
Lightweight browser-only metadata/static bundle.
- Multimodel desktop app with model selector,
Autorouting, collective mode, and agent mode. - Local chat, omni, image, vision, math, protein, 3D, and materials specialist model families.
- Curated desktop installer with built-in models plus one-click downloadable optional models.
- Benchmark sweeps and local graph generation for the model zoo.
- Hugging Face publishing helpers for models, installers, datasets, and downloadable model-store artifacts.
- Training pipelines for omni, frontier, native-image, and specialist lines.
- The latest finished full omni checkpoint in this workspace is
omni_collective_v7. omni_collective_v8remains the long-running local experiment.- The desktop line uses a curated built-in bundle plus an in-app Hugging Face model store instead of shipping the full legacy model zoo.
- The most recent desktop release is the V48 Frontier Edition.
python runtime_python/chat_web_app.pyWindows launchers:
runtime_python\launch_chat_web_supermix.bat
runtime_python\launch_chat_terminal_supermix.batpython source/chat_web_app.pypython source/supermix_multimodel_desktop_app.pypython source/training_monitor_gui.py --root .Open:
web_static/index.html
- Release page:
https://github.com/kai9987kai/Supermix/releases - Hugging Face desktop installer mirror:
https://huggingface.co/datasets/Kai9987kai/supermix-studio-desktop-installer - Model zoo:
https://huggingface.co/datasets/Kai9987kai/supermix-model-zoo
dist/SupermixStudioDesktop/SupermixStudioDesktop.exedist/installer/SupermixStudioDesktopSetup.exedist/installer/SupermixStudioDesktopReleaseSHA256.txt
The desktop app ships with a curated core bundle and can also download optional models from the in-app model store.
The current workspace includes code and/or packaged artifacts for:
v28v30
v31v32v33v34v35v39v40_benchmaxv41v46v47v48
v36v37v38
v1v2v3v4v5v6v7v8_previewv8in progress locally
math_equation_micro_v1science_vision_micro_v1protein_folding_micro_v1mattergen_micro_v1three_d_generation_micro_v1dcgan_mnist_modeldcgan_v2_in_progress
The latest finished full omni checkpoint in this workspace is omni_collective_v7.
The latest preview checkpoint is omni_collective_v8_preview.
omni_collective_v8 is still training locally and is not yet the published stable omni release.
Representative current results from the local benchmark inventory:
v40_benchmax:0.2433common benchmark scoreomni_collective_v8_preview:0.2167on the reduced preview benchmark sweepomni_collective_v7:0.1067on the current 6-benchmark local sweep
Specialist-only models remain in the graph inventory even when the common text benchmark is not the right evaluation fit.
Representative training and continuation scripts:
source/train_omni_collective_v2.py
source/train_omni_collective_v3.py
source/train_omni_collective_v4.py
source/train_omni_collective_v5.py
source/train_omni_collective_v6.py
source/train_omni_collective_v7.py
source/train_omni_collective_v8.py
source/train_math_equation_model.py
source/train_image_recognition_model.py
source/train_protein_folding_model.py
source/train_three_d_generation_model.py
source/train_mattergen_generation_model.py
source/benchmark_all_models_common.py
If you want the active experimental path, start in source/.
Primary desktop build helpers:
source/build_supermix_studio_desktop_exe.ps1
source/build_supermix_studio_desktop_installer.ps1
SupermixStudioDesktop.spec
installer/SupermixStudioDesktop.iss
The current desktop app supports:
- curated built-in bundle seeding
- in-app Hugging Face model-store downloads
- richer chat UI with drafts, context bank, compare bench, and dispatch preview
- richer training monitor with recovery posture, rescue guidance, and fleet spotlight summaries
Public Hugging Face repos used by this workspace include:
Kai9987kai/supermix-model-zooKai9987kai/supermix-studio-desktop-installer
Previously published model repos from this workspace include:
Kai9987kai/supermix-v33-frontierKai9987kai/supermix-omni-collective-v1Kai9987kai/supermix-v38-native-image-xlite-fp16Kai9987kai/supermix-v39-frontier-reasoning-plusKai9987kai/supermix-omni-collective-v2-frontierKai9987kai/supermix-math-equation-micro-v1Kai9987kai/supermix-omni-collective-v4-frontierKai9987kai/supermix-omni-collective-v7-frontier
Public dataset repos include:
Kai9987kai/supermix-conversation-datasetsKai9987kai/supermix-science-vision-dataset
- This repo is intentionally not a clean source-only model repo.
- Large local artifacts, checkpoints, and logs are often kept locally and mirrored selectively to releases or Hugging Face.
- The safest way to consume the desktop app is through the curated-core release plus the in-app model store.
MIT
::contentReference[oaicite:1]{index=1}