DeepSeek-V4-Flash-GGUF

DeepSeek V4 Flash — YaiLabs mixed GGUF representations

YaiLabs-derived GGUF files from deepseek-ai/DeepSeek-V4-Flash-DSpark. The original model was authored by DeepSeek; YaiLabs performed the conversion and mixed quantization using YVEX.

These exact bytes passed structural and tensor verification and the YVEX load → chat → unload lifecycle on the NVIDIA DGX Spark configuration documented below. The files are complete, unsharded representations; download only the representation you intend to use.

Downloads

Download and consume

With a current Hugging Face CLI, download the selected file and verify its SHA-256:

hf download yailabs/DeepSeek-V4-Flash-GGUF DeepSeek-V4-Flash-IQ2_XXS-Q2_K-Q8_0-v1.gguf SHA256SUMS --revision v1.0 --local-dir .
sha256sum DeepSeek-V4-Flash-IQ2_XXS-Q2_K-Q8_0-v1.gguf

Compare the result with SHA256SUMS. The v1.0 release tag names the verified release commit. For automation, resolve that tag once and pass its full Hub commit ID as --revision; do not rely on a floating main revision. For DeepSeek, substitute the other filename to select the second representation.

YVEX acquisition remains yvex model pull SOURCE; an externally downloaded file can be inspected through its existing local-path intake:

yvex model pull ./DeepSeek-V4-Flash-IQ2_XXS-Q2_K-Q8_0-v1.gguf --reference --dry-run --json

Acquisition is distinct from runtime preparation. A GGUF download alone does not create an admitted runtime binding/profile. The qualification used YVEX’s canonical binding preparation and an exact-SHA variant, followed by model load, hosted chat, model unload, and host shutdown. Use a YVEX revision containing the exact physical admission records (qualification revision below or a compatible successor). This release does not claim drop-in execution in every GGUF reader or third-party runtime.

Representation details

DeepSeek-V4-Flash-IQ2_XXS-Q2_K-Q8_0-v1.gguf

GGUF v3; architecture deepseek4; 1409 tensors; 1/1 files; 32-byte alignment.

Target expert gate/up tensors use IQ2_XXS and expert down tensors use Q2_K. This representation uses Q8_0 for quantizable nonexpert matrices; selected source-authored draft expert tensors retain MXFP4. Other tensor types remain as listed; a single conventional quantization label would be misleading.

Full tensor descriptors: CSV. Machine-readable identity, lineage, configuration and validation: release manifest.

Observed conversion/quantization/emission plus native roundtrip: 3441.594860 seconds, 2026-09-05T22:39:36.268070+00:00 to 2026-09-05T23:36:57.862154+00:00. Producing YVEX revision: 628d226aef82899fa1d95dd158d746a9938755d5; executable SHA-256: 4bde2b2dd34043bc3a5c2a0b8a99a5dfa1f06326bc0e8e0aa07343ac2a1262b4.

DeepSeek-V4-Flash-IQ2_XXS-Q2_K-MXFP4-v1.gguf

GGUF v3; architecture deepseek4; 1409 tensors; 1/1 files; 32-byte alignment.

Target expert gate/up tensors use IQ2_XXS and expert down tensors use Q2_K. This smaller representation preserves source MXFP4 for selected nonexpert matrices while retaining Q8_0 elsewhere. It is not an all-MXFP4 or all-Q2_K file. Other tensor types remain as listed; a single conventional quantization label would be misleading.

Full tensor descriptors: CSV. Machine-readable identity, lineage, configuration and validation: release manifest.

Observed conversion/quantization/emission plus native roundtrip: 3344.502410 seconds, 2026-09-05T22:40:00.901892+00:00 to 2026-09-05T23:35:45.403055+00:00. Producing YVEX revision: 628d226aef82899fa1d95dd158d746a9938755d5; executable SHA-256: 4bde2b2dd34043bc3a5c2a0b8a99a5dfa1f06326bc0e8e0aa07343ac2a1262b4.

Provenance and verification

Exact upstream: deepseek-ai/DeepSeek-V4-Flash-DSpark@62af8fffb2f7030cac4de2f0169f5b8d1101b646. Retained source shards were verified against the pinned file digests before transformation; source-file identities are included. Commands in the release manifests preserve effective arguments with portable placeholders replacing private local paths.

Each output passed full SHA-256 hashing, native emitter roundtrip, comparison with the sealed physical tensor plan, a GGUF metadata reader, and the pinned GGML reader at af97976c7810cdabb1863172f31c432dab767de7. Tensor manifests describe these exact output files.

The quantization policies are included under provenance/. The 450,892,654-byte calibration input is identified by SHA-256 in each release manifest and retained by YaiLabs; it is not bundled here. Bit-exact reconstruction requires that input as well as the pinned source and tooling.

Exact YVEX qualification

Validation used YVEX 1c7b7218e7264f3d3142fcfd046f85834809a30b (tree 2ce910bfd6096debc8d86dd76765f2bb8f1f9dd2) on NVIDIA DGX Spark / GB10: aarch64, 20 ARM CPU cores, 130,663,165,952 bytes RAM, Ubuntu 24.04.4 LTS, kernel 6.17.0-1021-nvidia, NVIDIA driver 580.159.03, CUDA toolkit 13.0.88.

The CUDA target-only lifecycle selected each artifact by its full SHA-256, loaded it, generated a nonempty text response to a short English greeting request, unloaded it, and stopped the host cleanly. Context capacity was 512; prefill chunk size 64; generation limit 32 tokens; temperature 0; top-p 1; reasoning disabled. Exact dates, outputs and lifecycle duration appear in each release manifest.

This is bounded compatibility evidence for these bytes and this configuration. It is not a general model-quality evaluation, a universal DGX Spark guarantee, an independent numerical-quality score, or a performance benchmark. Build durations include the observed machine load and scheduling. Exact peak working storage and separate historical download timings were not captured; sampled preparation-directory high-water marks are lower bounds.

License and attribution

The derivative weights remain governed by the upstream MIT license. See the unmodified LICENSE and NOTICE. YaiLabs claims conversion/quantization and validation authorship only, not original training or upstream endorsement.