Downloads · 30 days
10
17% of all-time downloads
coconut-qwen3-4b-ff
agurung/coconut-qwen3-4b-ff
coconut-qwen3-4b-ff is a machine learning model from agurung. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
This package was exported from Coconut checkpoint checkpoint32.
All-time downloads
58
Public
Parameters
4B
40.3 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors16.1 GB · 67%
At a glance
- Model type
- qwen3
- Access
- Public
- Created
- Mar 19, 2026
- Updated
- Oct 1, 2026
- SHA
- d6c2b0c5
Model card
From the Hugging Face model README
Coconut Export For Qwen/Qwen3-4B-Instruct-2507
This package was exported from Coconut checkpoint checkpoint_32.
What Is Included
- Standard HF model weights for
AutoModelForCausalLM.from_pretrained(...) - A tokenizer that includes the Coconut latent tokens used during training
- Metadata describing the latent-token ids and auxiliary Coconut state
Exact Coconut latent decoding
The export also includes latent_checkpoint.pt, which is the original Coconut checkpoint. That file preserves the wrapper state needed for exact latent decoding and checkpoint-faithful recovery.
Latent Tokens
<|start-latent|>->151669<|end-latent|>->151670<|latent|>->151671
Coconut-Only Auxiliary State
The original Coconut checkpoint contains wrapper-only keys in addition to the base LM:
embedding.weight
Standard Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("PATH_OR_HF_REPO")
tokenizer = AutoTokenizer.from_pretrained("PATH_OR_HF_REPO")
- Type
- qwen3
- Created
- Mar 19, 2026
- Updated
- Oct 1, 2026