models

Ready-made bitgpu models (zero conversion)

Each directory holds the two SMALL files bitgpu needs - manifest.json (the architecture contract + tensor byte ranges) and the aux file (lookup tables; for the ONNX dirs also norm gammas) - produced with tools/convert-gguf.py or tools/convert-onnx.py from the public exports. The big weights file is not here: it streams straight from the Hugging Face Hub, byte-for-byte unchanged, so nothing needs re-hosting and there is nothing to convert.

model	hidden / layers	weights	data file (streams from the Hub)
bonsai-1.7b-gguf	2048 / 28	~250 MB	prism-ml/Bonsai-1.7B-gguf (Bonsai-1.7B-Q1_0.gguf)
bonsai-4b-gguf	2560 / 36	~580 MB	prism-ml/Bonsai-4B-gguf (Bonsai-4B-Q1_0.gguf)
bonsai-8b-gguf	4096 / 36	~1.16 GB	prism-ml/Bonsai-8B-gguf (Bonsai-8B-Q1_0.gguf)
bonsai-27b-gguf	5120 / 64	~3.8 GB	prism-ml/Bonsai-27B-gguf (Bonsai-27B-Q1_0.gguf; Qwen3.5 hybrid, GGUF-only)
bonsai-1.7b	2048 / 28	~290 MB	onnx-community/Bonsai-1.7B-ONNX
bonsai-4b	2560 / 36	~660 MB	onnx-community/Bonsai-4B-ONNX
bonsai-8b	4096 / 36	~1.3 GB	onnx-community/Bonsai-8B-ONNX

The GGUF and ONNX containers carry the same weights (the sign-bit streams are bit-identical; verified on Bonsai-8B): pick by hosting preference. GGUF is the primary path - the files are a little smaller on the wire and their manifests synthesize RoPE at load (see docs/FORMAT.md); the ONNX dirs remain fully supported.

Use them from any static host or a CDN over this repo (pin a tag for immutable caching):

import { createEngine } from 'bitgpu'
import { createChat } from 'bitgpu/chat'

const REPO = 'https://cdn.jsdelivr.net/gh/stfurkan/bitgpu@v0.19.1/models/bonsai-1.7b-gguf'
const TOK = 'https://huggingface.co/onnx-community/Bonsai-1.7B-ONNX/resolve/main'

const engine = await createEngine({
  manifestUrl: `${REPO}/manifest.json`,
  auxUrl: `${REPO}/Bonsai-1.7B-Q1_0.aux.bin`,
  dataUrl: 'https://huggingface.co/prism-ml/Bonsai-1.7B-gguf/resolve/main/Bonsai-1.7B-Q1_0.gguf',
})
// or the ONNX flavor of the same weights:
//   REPO = '.../models/bonsai-1.7b'  auxUrl: `${REPO}/bonsai.aux.bin` (model_q1.aux.bin for 4B/8B)
//   dataUrl: `${TOK}/onnx/model_q1.onnx_data`
const chat = await createChat(engine, {
  tokenizerJsonUrl: `${TOK}/tokenizer.json`, // tokenizers live in the ONNX/source repos either way
  tokenizerConfigUrl: `${TOK}/tokenizer_config.json`,
})

Or copy a directory into your app's static assets and own the files (they are just static files - your repo becomes their source control). Two variants:

• copy manifest + aux only and keep dataUrl pointed at the Hub: two small first-party files, weights still stream from HF;
• also download the data file (and tokenizer files) into the same directory for a fully first-party, offline-capable setup with no third-party requests - then the single-URL form works, since the manifest names its data/aux files and the engine resolves them relative to the directory: createEngine({ modelUrl: '/models/bonsai-1.7b' }).

For GGUF models these files are a convenience, not a requirement: bitgpu/gguf's fromGguf(ggufUrl) builds the same manifest in the browser from the GGUF header alone (gated deep-equal to these committed ones). The committed manifests remain the recommended path when they exist - smaller transfer, better caching.

These files are intentionally not shipped in the npm package: bitgpu stays a lean, model-neutral engine - hotlink the pinned CDN URLs above, or copy the files and own them. They are static and versioned with the engine; regenerating them from the exports reproduces them byte-for-byte (python tools/convert-gguf.py --gguf <file> / convert-onnx.py --model <dir>, full pipeline in tools/README.md).