Qwen3.5-0.8B-OCPI

Qwen3.5-0.8B-OCPI

Finetune packaging of Qwen/Qwen3.5-0.8B as a full-weight OCPI operator field (not digests-only).

Method

This is an operator-field finetune packaging (method OCPI_FIELD in adapter_config.json): full base weights re-encoded for carrier/runtime use with projection ledger fidelity proofs. It is not a LoRA adapter.

Files

• model.safetensors — full f16 weights (export from OCPI field)
• config.json, tokenizer files — inference config
• adapter_config.json — hub-facing finetune metadata
• Optional companion: ocpi_field/ (binary lattice + events)

Multi-arch runtime

Standalone Rust binary ocpi-qwen is published for:

• Windows: x86_64, i686, ARM64
• Linux: x86_64, i686, aarch64
• macOS: x86_64 + Apple Silicon with Metal feature

ocpi-qwen verify --field ocpi_field --lattice
ocpi-qwen chat --field ocpi_field --prompt "Hello"
ocpi-qwen export-hf --field ocpi_field --out ./

Transformers load

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained(".", trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    ".", trust_remote_code=True, torch_dtype=torch.float16
)
messages = [{"role": "user", "content": "Hello! Who are you?"}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=48)
print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Citation

Base model: Qwen Team — Qwen3.5. OCPI field packaging: local PoC (ocpi.poc.qwen08b.v1).