---
title: DeepSeek-V4-Flash-0731-GGUF-SpectraX
canonical_url: "https://www.modelscope.ai/models/SenkuLabs/DeepSeek-V4-Flash-0731-GGUF-SpectraX"
md_url: "https://www.modelscope.ai/models/SenkuLabs/DeepSeek-V4-Flash-0731-GGUF-SpectraX.md"
repository: SenkuLabs/DeepSeek-V4-Flash-0731-GGUF-SpectraX
last_updated: 2026-08-08
downloads: 233
stars: 0
---

# DeepSeek-V4-Flash-0731-GGUF-SpectraX

> DeepSeek-V4-Flash-0731-GGUF-SpectraX - An open-source model by SenkuLabs on ModelScope. SpectraX/EasyDeL-ready conversions of every DeepSeek-V4-Flash-0731 Unsloth GGUF quantization, with immutable provenance and per-stage timing receipts.

- **Repository**: SenkuLabs/DeepSeek-V4-Flash-0731-GGUF-SpectraX
- **Downloads**: 233
- **Stars**: 0
- **Last updated**: 2026-08-08

Source: https://www.modelscope.ai/models/SenkuLabs/DeepSeek-V4-Flash-0731-GGUF-SpectraX

---

# DeepSeek-V4-Flash-0731-GGUF-SpectraX

SpectraX/EasyDeL-ready conversions of every quantization in `unsloth/DeepSeek-V4-Flash-0731-GGUF` on ModelScope.

## Source

- Source repository: `unsloth/DeepSeek-V4-Flash-0731-GGUF`
- Source revision: `master` (resolved and recorded per worker before conversion)
- Source inventory: 13 quantization directories, 52 GGUF shards, 1,514,912,884,655 bytes
- License: MIT
- Architecture: DeepSeek V4 (`deepseek4`)

## Publication layout

`master` contains only the immutable repository contract, source inventory, and verified merge index.
Each quantization is converted and uploaded on an isolated worker branch named:

```text
worker/<quantization-lowercase>-spectrax-v1
```

A branch is eligible for merge only after all of the following pass:

1. Every source shard is pinned by path, byte size, and SHA-256.
2. Every source tensor is consumed exactly once.
3. Every SpectraX target leaf is produced exactly once.
4. Per-stage timing receipts cover download, checksum, GGUF decode, tensor transform, quantization/packing, local write, upload, and remote readback.
5. Local and remote file counts, byte sizes, SHA-256 values, and tree SHA-256 match.
6. The converted checkpoint contains native SpectraX/EasyDeL leaves and no unresolved external GGUF tensors.

## Target quantization policy

The first publication profile uses kernel-aligned affine packing while preserving the source family as a separate revision:

| Source family | SpectraX storage |
|---|---:|
| IQ1 | affine INT2 |
| IQ2 / Q2 | affine INT2 |
| IQ3 / Q3 | affine INT4 |
| IQ4 / Q4 | affine INT4 |
| Q8 | affine INT8 |

This avoids silently collapsing all source variants to INT2. Every branch records the exact source quantization, target bit width, group size, packing axis, source revision, converter commit, and numerical validation statistics.

## CPU conversion

Workers default to all logical CPUs reported by `os.cpu_count()` and set BLAS/OpenMP inner thread counts to one, preventing nested oversubscription. The worker count, CPU topology, process IDs, per-tensor throughput, aggregate bytes/second, and wall-clock duration are recorded in JSONL receipts.

## Status

The repository baseline is created. Quantization branches are populated only by the resumable `scripts/gguf_spectrax` pipeline in the EasyDeL monorepo.