Qwen3.5 8-layer tiny random chat checkpoint

Random weights for software tests, not a trained assistant. Text only; no image/audio/video support. Original Qwen3.5-4B tokenizer, full vocabulary, chat/thinking/tool template and special tokens. Width 64, MLP 128, 8 layers, 16,227,664 parameters, bf16 safetensors with tied embeddings. Keeps Qwen3.5's three linear-attention layers followed by one full-attention layer. RoPE sections are resized to [2, 1, 1] for head dimension 32 and rotary fraction 0.25. CPU operation needs no FlashAttention or FLA installation; optimized GPU kernels are optional.

from transformers import AutoModelForCausalLM, AutoTokenizer

path = "wassname/qwen3.5-8lyr-tiny-random"
tok = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path, attn_implementation="eager")
inputs = tok.apply_chat_template(
    [{"role": "user", "content": "Hello!"}],
    add_generation_prompt=True, enable_thinking=False,
    return_tensors="pt", return_dict=True,
)
reply = model.generate(**inputs, max_new_tokens=8, do_sample=False)

For attention weights, load with attn_implementation="eager" and pass output_attentions=True. Only full-attention layers return maps; linear-attention layers use recurrent states, not attention maps. For eight layers the returned maps correspond to zero-based layers 3 and 7, not all eight layers. SDPA supports generation but does not return these maps in the tested Transformers version.

Built from Qwen/Qwen3.5-4B@851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a with seed 42; no pretrained weights copied. Build: uv run python -m lucid3.tiny_qwen --out <new-directory> --layers 8 --seed 42. Source tokenizer/config/LICENSE must be cached locally. See build.json and verification.json. Checked saved-model forward/backward, cached decoding, padded batches, generation and template preservation. Use --layers 32 for LUCID tests that access layers 16 and 30. Eight layers cannot supply those layers. Random outputs are meaningless; successful generation does not establish chat or tool-use ability.

Downloads last month
27
Safetensors
Model size
16.2M params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for wassname/qwen3.5-8lyr-tiny-random

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(1,449)
this model