BotPsy 0.8B WebGPU
run the model right in your browser
2 min readupdated
This is a build of the compact qvac/BotPsy-0.8B language model (0.8B parameters), packaged to run directly in the browser via WebGPU (a web API for hardware-accelerated graphics and compute). Unlike standard repositories using Transformers or GGUF formats, the model comes in custom bundles with manifests and shards (split weight files) that download and cache into the browser's local OPFS storage. This repository is useful for web developers integrating the BotPsy chat app who want to run client-side inference without a server backend.
Key features
- Three text-weight quantization options:
Q8_0, the lighterQ4_K_M, and amixedbuild (attention projections in Q8, the rest of the model in Q4_K_M). - A separate
botpsy-vision-bf16-f32package containing a 12-layer vision encoder (vision tower) for handling images. - SHA-256 checksum verification for every shard and asset before installing into OPFS.
- A ready-to-use directory structure with tokenizers, chat templates, and assets for plugging straight into the browser runtime.
What you can do
- Connect the model to the BotPsy web app for local in-browser chat inference with hash verification and OPFS caching.
- Pick the right balance between download speed and response accuracy using the prebuilt
Q4_K_M,Q8_0, ormixedpackages. - Host your own mirror of the model files and override the base download URL during the frontend build via the
VITE_MODEL_BASE_URLenvironment variable.
Example
# Загрузка манифеста выбранного пакета в браузерном рантайме:
[https://huggingface.co/leventkoldas/botpsy-0.8b-webgpu/resolve/main/botpsy-0.8b-q4_k_m/manifest.json](https://huggingface.co/leventkoldas/botpsy-0.8b-webgpu/resolve/main/botpsy-0.8b-q4_k_m/manifest.json)
# Переопределение базового URL при сборке приложения:
VITE_MODEL_BASE_URL=[https://huggingface.co/leventkoldas/botpsy-0.8b-webgpu/resolve/main](https://huggingface.co/leventkoldas/botpsy-0.8b-webgpu/resolve/main)Read more
DetailsBase model: qvac/BotPsy-0.8B (revision 24561e9681643eaf807fc28b0875210cb17f89c1)
Weights and architecture license: Apache-2.0
Format: custom packages with manifest.json and shards for the WebGPU runtime (not directly compatible with Transformers or GGUF)
Available directories: botpsy-0.8b-q8_0, botpsy-0.8b-q4_k_m, botpsy-0.8b-mixed, botpsy-vision-bf16-f32










Comments