Qwen3.8-Flash-Next-GGUF

Qwen3.8-Flash-Next GGUF

Support my work

I work on making large language models practical on hardware they were never really designed to fit on — through mixed quantization, inference optimization, custom kernels, and serving experiments.

While much of the development happens on local hardware, calibration, profiling, and large-scale validation often require expensive on-demand GPUs.

Contributions help pay for that compute, storage, and testing infrastructure, so I can keep experimenting and publishing the results openly.

<a href="https://www.buymeacoffee.com/baekpica" target="_blank"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me a Coffee" style="height: 60px !important;width: 217px !important;"></a> <a href="https://github.com/sponsors/Baekpica" target="_blank"><img src="https://img.shields.io/badge/Sponsor-EA4AAA?style=for-the-badge&logo=githubsponsors&logoColor=white" alt="Sponsor Baekpica on GitHub" style="height: 60px !important;width: 217px !important;"></a>

GGUF conversions of Qwen/Qwen3.8-Flash-Next, pinned to source revision f5d08274bafd880402bd16f5e3e6c514136ec06c.

Compatibility status

These files use a custom qwen4exp GGUF architecture prepared for the dfm branch of Baekpica/ds4, with DGX Spark/GB10 as the deployment target. Runtime support is still under implementation and validation. Do not assume compatibility with upstream llama.cpp or other GGUF runtimes unless they explicitly support this architecture and tensor schema.

Available variants

A separately tuned mixed-quant release will be published only after calibration, H200 quality checks, and ds4 runtime validation.

BF16 verification

The BF16 release contains 1,756 GGUF tensors in 12 shards, totaling 360,011,029,056 bytes. It was checked against all 1,658 tensors from the pinned source revision:

• all 359,999,963,128 source payload bytes compared exactly;
• zero payload mismatches;
• fused expert gate/up tensors were split semantically without numerical conversion;
• routed expert down projections were split losslessly into the main 512-column region and the 128-column tail;
• GGUF shard metadata, tensor offsets, shapes, types, and Qwen Community License metadata were validated;
• per-shard SHA-256 checksums are included alongside the files.

The source snapshot itself was also checksum-verified with hf cache verify before conversion.

Q8_0 verification

The Q8_0 release contains 1,756 GGUF tensors in 7 shards, totaling 192,201,208,384 bytes. Its audited tensor distribution is 806 Q8_0, 363 BF16, 584 F32, and 3 I64 tensors. GGUF metadata, shard numbering, tensor names, shapes, offsets, declared payload ends, and types were checked against the same exhaustive pinned source map with zero structural errors.

Per-shard SHA-256 checksums are included in Q8_0/SHA256SUMS. All seven public Hugging Face LFS object hashes and remote byte sizes were also compared with the local artifacts after upload and matched exactly.

Architecture notes

This conversion retains the complete multimodal and speculative-decoding topology: 48 text layers, 36 gated-delta layers, 12 full-attention layers, 512 routed experts with top-10 routing, shared experts, four hyper-connection streams, a 51.2B-parameter PLE n-gram table, the vision tower, and the MTP layer.

License

Use of these converted weights is governed by the original Qwen Community License 1.0. The exact upstream license file is included in this repository. No Apache-2.0 license is claimed for these weights.

Reproducibility

Conversion, verification, calibration, mixed-quant recipe, and ds4 runtime materials are being prepared for publication with the validated mixed-quant handoff. Until runtime validation is complete, the BF16 and Q8_0 files should be treated as conversion artifacts rather than a ready-to-run general-purpose release.