tinywmcraft is an open-weight model from Fedor, released under MIT License. Its published files total 396.8 MB.
Checkpoints of tinywmcraft, a 13.9M-parameter pixel world model for Craftax-Classic trained from scratch in 10 minutes on one A100.
Model Card
By Fedor, published under mit, revision 05aa04867ed0.
Checkpoints of tinywmcraft, a 13.9M-parameter pixel world model for Craftax-Classic trained from scratch in 10 minutes on one A100. Each model.npz holds the tokenizer, the flow-matching world model with reward and termination heads, and JSON metadata with the training config and time. Recipe on main: hybrid tile tokenizer, Pallas kernels, Muon, reward and termination heads. Trained cold on one Colab A100-SXM4-40GB on 2026-10-15, scored by the frozen eval.py (2048 windows, 8 context frames, 16 generated, 8 Euler steps). Mean gFVD 0.00147 ± 0.00019. Each folder has the ledgerrow.jsonl written by eval.py. An earlier research recipe (branch normuon): gFVD 0.00125, AF 0.920, seed 0. Lower gFVD…
Read Fedor's full model card
tinywmcraft checkpoints
Checkpoints of tinywmcraft, a 13.9M-parameter pixel world model for
Craftax-Classic trained from scratch in 10 minutes on one A100. Each model.npz holds the tokenizer, the
flow-matching world model with reward and termination heads, and JSON metadata with the training config and time.
Record 1
Recipe on main: hybrid tile tokenizer, Pallas kernels, Muon, reward and termination heads.
Trained cold on one Colab A100-SXM4-40GB on 2026-10-15, scored by the frozen eval.py (2048 windows, 8 context frames, 16 generated, 8 Euler steps).
| folder | seed | gFVD | AF | reward precision | train s |
|---|---|---|---|---|---|
record1_main_s0 |
0 | 0.00135 | 0.927 | 0.899 | 578 |
record1_main_s1 |
1 | 0.00169 | 0.937 | 0.913 | 581 |
record1_main_s2 |
2 | 0.00137 | 0.924 | 0.878 | 579 |
Mean gFVD 0.00147 ± 0.00019. Each folder has the ledger_row.jsonl written by eval.py.
curl -L --fail --create-dirs -o runs/record1/model.npz https://huggingface.co/maharishiva/tinywmcraft/resolve/main/record1_main_s0/model.npz
python play_wm.py --model runs/record1/model.npz
normuon_head_bridge_s0
An earlier research recipe (branch normuon): gFVD 0.00125, AF 0.920, seed 0. Lower gFVD, lower action following and reward precision; not a record. Play it from that branch.
Identity and Version
- Repository
- maharishiva/tinywmcraft
- Publisher
- Fedor
- Task
- Not stated by the source
- Modality
- Other
- Library
- jax
- Parameters
- Not stated by the source
- Languages
- jax
- Revision
- 05aa04867ed0c1348e775d00eaafa77cf7b5ddf9
- First published
- 2026-10-13
- Last updated
- 2026-10-16
Files and Weights
16 files, 396.8 MB in total. The weights are 8 files totalling 396.8 MB in npz, onnx.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| normuon_head_bridge_s0/model.npz | Weights | 55.5 MB | b4d7b82ba698 |
| record1_main_s0/model.npz | Weights | 55.5 MB | fe3227034c90 |
| record1_main_s1/model.npz | Weights | 55.5 MB | d748fbebee3e |
| record1_main_s2/model.npz | Weights | 55.5 MB | cc9cec048798 |
| train_tiny_s0/model.npz | Weights | 50.4 MB | b904c13845aa |
| train_tiny_s1/model.npz | Weights | 50.4 MB | 8793cc0045d1 |
| train_tiny_s2/model.npz | Weights | 50.4 MB | 7d1b2b5336cc |
| web/play.onnx | Weights | 23.6 MB | bcbbe39548b8 |
| README.md | Documentation | 1.5 KB | — |
| record1_main_s0/ledger_row.jsonl | Other | 540 B | — |
| record1_main_s1/ledger_row.jsonl | Other | 540 B | — |
| record1_main_s2/ledger_row.jsonl | Other | 539 B | — |
| train_tiny_s0/ledger_row.jsonl | Other | 616 B | — |
| train_tiny_s1/ledger_row.jsonl | Other | 618 B | — |
| train_tiny_s2/ledger_row.jsonl | Other | 619 B | — |
| .gitattributes | Repository | 1.5 KB | — |
License and Download
- License
- mit
- Access
- Open weights, no gate
- Download size
- 396.8 MB
Released by Fedor through its official repository on Hugging Face. Read the license.
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 396.8 MB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About tinywmcraft
Can I use tinywmcraft commercially?
Yes. tinywmcraft is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.