Total runs:
2.5K
Run Growth:
1.7K
Growth Rate:
68.68%
Sponsored by
APIMart
eulogik / nanowhale-100m-DIME
Model's Last Updated: July 19 2026
text-generation
Introduction of nanowhale-100m-DIME
Model Details of nanowhale-100m-DIME
๐ nanowhale-100m-DIME
Pure-NumPy reference implementation of the DeepSeek-V4-Tiny architecture (MLA + Hyper-Connections + MoE).
Maintained by eulogik โ building DIME , the disk-streaming Mixture Engine for running 284B-class MoE models on 16 GB Macs.
No PyTorch. No GPU. No CUDA. Just
numpy
+
safetensors
. And it matches the
official PyTorch reference
exactly
โ cosine similarity =
1.000
on logits,
verified across 1โ9 tokens.
This is the reference engine behind DIME (Disk-streaming Mixture Engine) โ the runtime that runs 284B-class MoE models on a 16 GB Mac Mini by streaming experts from disk.
โก Run a 110M-parameter MoE model in ~10 lines, in the browser, on a phone, or on a toaster. The weights are ~441 MB and load with zero ML frameworks.
๐ Quickstart
pip install numpy safetensors
from huggingface_hub import snapshot_download
from nanowhale_dime import NanowhaleDIME
model_dir = snapshot_download("cmpatino/nanowhale-100m")
model = NanowhaleDIME(model_dir)
print(model.generate("The future of efficient inference is", max_new_tokens=32))
Or just open the Colab notebook โ zero setup.
๐ง What's inside
nanowhale-100m
is a 110M-parameter
DeepSeek-V4-Tiny
that distills the key
ideas of DeepSeek-V4 into a tiny, hackable package:
| Component | Detail |
|---|---|
| Layers | 8 |
| Attention | Multi-head Latent Attention (MLA) with RoPE |
| Experts | 4 routed + 1 shared, top-2 routing |
| Hyper-Connections |
hc_mult=4
with Sinkhorn-iterating mixer
|
| MTP | Multi-Token Prediction head |
| Params | ~110M (float32, ~441 MB safetensors) |
All of it is reimplemented in readable NumPy in a single file
(
nanowhale_dime.py
) that you can read end-to-end in one
sitting. Great for learning how modern MoE + MLA models actually work under the
hood.
โ Why "bit-exact"?
We compared our NumPy forward pass against the official PyTorch reference model on logits for sequences of length 1, 4, and 9 tokens:
tokens=1 cos=1.000
tokens=4 cos=1.000
tokens=9 cos=1.000
The implementation faithfully reproduces:
- RMSNorm with the exact variance epsilon
- MLA query/key de-rotation (inverse RoPE on attention output)
- The Softplus + gate-bias router normalization
- Hyper-Connection pre/post mixing with Sinkhorn-balanced combination matrix
- Shared + routed expert summation with top-2 scaling
๐ What is DIME?
DIME is an MoE inference runtime for Apple Silicon that breaks the "must fit in RAM" rule. It lazily loads expert weights from local or networked NVMe and keeps a frequency-promoted hot tier of the experts you actually use (the same trick ktransformers uses to run 236B models on consumer hardware).
nanowhale-100m-DIME
is the fully self-contained, verified reference. Once the
math is proven at 110M params, the same engine scales to
DeepSeek-V4-Flash-class (284B) models
โ streamed, not loaded.
Follow eulogik for the 284B-on-a-Mac-Mini release. ๐
๐ Original model
Weights and tokenizer are from
cmpatino/nanowhale-100m
.
This repo contains only the NumPy inference engine + Colab + this card.
๐ Maintained by eulogik
We build DIME (Disk-streaming Mixture Engine) โ MoE inference for Apple Silicon that runs models far larger than RAM by streaming experts from disk. Follow @eulogik and star eulogik/dime for the 284B-on-a-Mac-Mini release. ๐
๐ License
Apache-2.0 (weights inherit their original license).
Runs of eulogik nanowhale-100m-DIME on huggingface.co
18
Total runs
0
24-hour runs
0
3-day runs
0
7-day runs
4
30-day runs
More Information About nanowhale-100m-DIME huggingface.co Model
Url of nanowhale-100m-DIME
nanowhale-100m-DIME huggingface.co Url
Provider of nanowhale-100m-DIME huggingface.co
Other API from eulogik
Updated:September 23 2026
Total runs:
813
Run Growth:
813
Growth Rate:
100.00%
Total runs:
631
Run Growth:
251
Growth Rate:
39.78%
Total runs:
455
Run Growth:
455
Growth Rate:
100.00%
Total runs:
424
Run Growth:
307
Growth Rate:
72.41%
Total runs:
309
Run Growth:
254
Growth Rate:
82.20%
Total runs:
278
Run Growth:
278
Growth Rate:
100.00%
Total runs:
241
Run Growth:
241
Growth Rate:
100.00%
Total runs:
206
Run Growth:
70
Growth Rate:
33.98%
Total runs:
72
Run Growth:
72
Growth Rate:
100.00%
Total runs:
55
Run Growth:
55
Growth Rate:
100.00%
Total runs:
41
Run Growth:
-160
Growth Rate:
-390.24%
Total runs:
34
Run Growth:
-16
Growth Rate:
-47.06%
Total runs:
30
Run Growth:
-13
Growth Rate:
-43.33%
Total runs:
30
Run Growth:
21
Growth Rate:
70.00%
Total runs:
27
Run Growth:
1
Growth Rate:
3.70%
Total runs:
24
Run Growth:
24
Growth Rate:
100.00%
Total runs:
14
Run Growth:
-101
Growth Rate:
-721.43%
Total runs:
7
Run Growth:
7
Growth Rate:
100.00%
Total runs:
0
Run Growth:
0
Growth Rate:
0.00%
Total runs:
0
Run Growth:
0
Growth Rate:
0.00%
Total runs:
0
Run Growth:
0
Growth Rate:
0.00%