naklitechie/swage-models

swage models

Browser artefacts for swage — Program-as-Weights (PAW, arXiv 2607.02512) in one tab. Source and contract: https://github.com/NakliTechie/swage (SPEC.md §6).

• interpreter/interpreter-fp16.onnx (+ .onnx.data, 1.5 GB) — Qwen3-0.6B exported to ONNX with the LoRA adapter as graph inputs (lora_A.<module>, lora_B.<module>, 14 fp16 tensors) and logits at requested positions; fused for ORT-web WebGPU. Derivative of Qwen/Qwen3-0.6B (Apache-2.0).
• encoder/encoder-q8.onnx + encoder-q8.part0..3.data (4.56 GB total, parts ≤ 1.5 GB) — the published PAW compiler (Qwen3-4B architecture) as an encoder: 28 depth-ratio taps at the last 64 positions, MatMulNBits 8-bit block 64. External data spans the four part files (per-tensor location + offset); ORT-web takes them as an externalData array.
• mapper/mapper.onnx (+ .onnx.data, 326 MB) — the PAW LoRA mapper in fp16; its outputs are exactly the interpreter's LoRA inputs.
• demo/<id>/ — three PAW programs (PEFT LoRA adapters, rank 64) compiled with the published PAW compiler via program-as-weights-server; same layout as programasweights/paw-programs.

Parity receipts (vs MLX / torch references) are in the source repo under export/receipts/.

Licences: Qwen3-0.6B derivatives Apache-2.0. The PAW compiler and mapper weights (programasweights/paw-4b-qwen3-0.6b) carry no licence file upstream; these re-quantised copies are published for research use while terms are confirmed with the authors. Prompt text inside the programs is (c) the PAW authors, CC BY 4.0.