Noema-Overfit

NoemaAI-labs/Noema-Overfit

Noema-Overfit is a machine learning model from NoemaAI-labs. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.

Paged MoE model bundles for Noema's Overfit expert-paging runtime. Each subfolder is one .noema-paged package: a resident.gguf (all non-expert weights, always loaded) plus experts-.bin page files that are streamed and…

Downloads · 30 days

371

27% of all-time downloads

All-time downloads

1.4K

Public

Repo size

247 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.bin232 GB · 94%

At a glance

License: other
Access: Public
Created: Jul 18, 2026
Updated: Aug 23, 2026
SHA: db1bf335

Base models

Model card

From the Hugging Face model README

Noema Overfit

Paged MoE model bundles for Noema's Overfit expert-paging runtime. Each subfolder is one .noema-paged package: a resident.gguf (all non-expert weights, always loaded) plus experts-*.bin page files that are streamed and evicted on demand, described by manifest.json.

These are runtime-specific packages, not standalone GGUF files. Each manifest.json declares its required Noema native contract: the Qwen and Gemma packages below require contract v3, while DeepSeek V4 requires contract v4.

Download the complete .noema-paged folder and preserve its layout. In Noema, add the package to Stored, select Automatic under Overfit (Paged Experts), and run the canary test before relying on it. A fast local SSD is strongly recommended. Expert paging increases runnable capacity; it does not turn storage into RAM or guarantee interactive speed.

Models

`DeepSeek-V4-Flash-0731-UD-IQ4_NL-00001-of-00004.noema-paged/`


Base model	`deepseek-ai/DeepSeek-V4-Flash-0731`
Source GGUF	`unsloth/DeepSeek-V4-Flash-0731-GGUF/UD-IQ4_NL` — 4 shards, 136.66 GB
Upstream revisions	Base `7872f01b1d1fe23eabc4c98b48bffcef5a386062` · GGUF `fbbb5b93fb787c21338159b0af3318bb3f4d9768`
Architecture	`deepseek4` (256 routed experts, 6 active, 43 MoE layers)
Quantization	`UD-IQ4_NL`
Resident weights	`resident.gguf` — 7.81 GB
Expert pages	`experts-000.bin` … `experts-007.bin` — 128.85 GB total
Manifest	Format v1 · native contract v4 · 33,024 records · 16,384-byte alignment
Source SHA-256	`e53b8a27242a271af2ebee6171763e07913fb54dbacb3745afa76e0c50c062ff`
Package fingerprint	`dcd6721b6c851be88ab192280dc9defc855e3f7f31e69b8db3bdcd8024727b12`
Manifest SHA-256	`d9a22388dca29a6abb4c296283f60f7a8b94b88b91d38e1107de05df5d46a32c`
License	MIT (upstream model)

`gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.noema-paged/`


Base model	`google/gemma-4-26B-A4B-it-qat-q4_0-unquantized`
Source GGUF	`unsloth/gemma-4-26B-A4B-it-qat-GGUF` — `UD-Q4_K_XL` (14.25 GB)
Architecture	`gemma4` (128 experts, 8 active, 30 MoE layers, fused gate/up)
Resident weights	`resident.gguf` — 1.40 GB
Expert pages	`experts-000.bin` — 12.96 GB
Alignment	16384 bytes

`Qwen3.6-35B-A3B-UD-Q4_K_M.noema-paged/`


Base model	`Qwen/Qwen3.6-35B-A3B`
Architecture	`qwen35moe` (256 experts, 8 active, 40 MoE layers)
Source GGUF	`Qwen3.6-35B-A3B-UD-Q4_K_M.gguf` (22.13 GB)
Resident weights	`resident.gguf` — 2.57 GB
Expert pages	`experts-000.bin`, `experts-001.bin` — 19.57 GB total
Alignment	16384 bytes

`Qwen3.5-122B-A10B-Q4_K_M.noema-paged/`


Architecture	`qwen35moe` (256 experts, 8 active, 48 MoE layers)
Source GGUF	`Qwen3.5-122B-A10B-Q4_K_M` (2 shards, 74.2 GB)
Resident weights	`resident.gguf` — 4.0 GB
Expert pages	`experts-000.bin` … `experts-004.bin` — 65 GB
Alignment	16384 bytes

Packages are generated with Noema's paged-model conversion tooling. File sizes, source fingerprints, and SHA-256 integrity hashes are recorded in each package's manifest.json.

Licenses

Noema's packaging does not replace the upstream model licenses. Use each bundle according to the terms attached to its base model:

DeepSeek V4 Flash 0731: MIT.
Qwen3.5 and Qwen3.6 bundles: Apache-2.0.
Gemma 4 bundle: Apache-2.0.

License: other
Created: Jul 18, 2026
Updated: Aug 23, 2026