---

license: other

license_name: qwen-community-1.0

license_link: https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/f5d08274bafd880402bd16f5e3e6c514136ec06c/LICENSE

base_model: Qwen/Qwen3.8-Flash-Next

base_model_relation: quantized

pipeline_tag: image-text-to-text

tags:

- gguf

- mixed-quant

- qwen4exp

- qwen3.8-flash-next

- dgx-spark

- ds4

---

Qwen3.8-Flash-Next Mixed-Quant 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>

> Runtime work starts with the SSD-offload variant: ds4 implementation and DGX Spark validation are being carried out first against Baekpica/Qwen3.8-Flash-Next-Mixed-Quant-SSD-PLE-GGUF, the SSD-offload track for this model family.

> Q2_K artifact available: the four-shard no-imatrix MQ-NG5-Q2K-Q4-SPLIT-E4 GGUF is published below. Its local structure/full-file hashes and remote LFS hashes are verified. Model quality, ds4 runtime behavior, and DGX Spark serving remain provisional until the later validation gates are complete. The calibrated MQ-NG5-Q4-SPLIT-E5 recipe remains published for review, but its weight build is paused while the higher-priority SSD-PLE Q5/Q6 track is produced.

This repository hosts DGX Spark/GB10-oriented mixed-precision GGUF conversions of Qwen/Qwen3.8-Flash-Next, pinned to source revision f5d08274bafd880402bd16f5e3e6c514136ec06c.

The planned runtime is the dfm branch of Baekpica/ds4, initially grounded at ds4 tag v0.6.3-dfm. This is a custom qwen4exp GGUF schema; compatibility with upstream llama.cpp or other GGUF runtimes is not implied.

Release tracks

Two tracks keep the calibration dependency and precision trade-off explicit:

| Candidate | Interior routed gate/up | Calibration dependency | Artifact size | Status | Conservative Spark headroom |

|---|---:|---:|---:|---:|---:|

| MQ-NG5-Q2K-Q4-SPLIT-E4 | Q2_K | none; standard weight-only quantization | 98.50165 GB / 91.7368 GiB actual | published; integrity verified | 4.10 GiB projected |

| MQ-NG5-Q4-SPLIT-E5 | IQ2_XXS on layers 3–45 | merged official-model activation importance matrix | 93.85 GB / 87.40 GiB projected | calibration passed; recipe retained, weight build paused | 8.43 GiB projected |

The Q2_K candidate was produced first. “No imatrix” means its Q2_K/Q4_K blocks are quantized directly from the pinned BF16 weights; it does not waive H200-quality, runtime, or DGX Spark validation.

Published Q2_K files

The complete artifact is in MQ-NG5-Q2K-Q4-SPLIT-E4/. It contains four GGUF shards, SHA256SUMS, the quantization recipe, artifact manifest, and strict structure-verification report.

• Total: 98,501,648,608 bytes / 91.7368 GiB.
• Structure: 1,756 tensors, zero verifier errors.
• Type counts: 88 Q2_K, 56 Q4_K, 48 Q4_0, 128 Q5_1, 486 Q8_0, 363 BF16, 584 F32, and 3 I64 tensors.
• Integrity: independent local full-file SHA-256 passed; all four remote LFS SHA-256 values and byte counts match the local manifest.

Calibrated E5 decision

Three official-model H200 passes contributed 1,510,400 processed tokens and 724,992,000 routed selections across a primary instruction/reasoning mix and an independent expanded multilingual/chat/science/table/code/safety/web corpus. The initial calibrated E4 scope had no zero cells and p1 3,167.97, but nine cells in layer 2 remained below the locked 512-route minimum; its least-used expert was observed only twice.

The release recipe therefore promotes layer 2 gate/up to Q4_K instead of forcing an under-supported IQ2_XXS quantization. Over the final IQ2 scope, layers 3–45, the minimum is 545 routes, p1 is 1,193.93 routes, and zero-coverage cells are 0. All pass inputs, corpus hashes, overlap accounting, accumulator/imatrix hashes, and thresholds are recorded in the MQ-NG5-Q4-SPLIT-E5 recipe. These route floors are operational calibration criteria, not a claim of downstream model quality.

| Model region | Q2_K no-imatrix track | Calibrated IQ2_XXS track | Reason |

|---|---:|---:|---|

| 51.2B-parameter PLE n-gram table (128 shards) | Q5_1 | Q5_1 | Dominant residency cost; higher precision than the most aggressive routed experts while preserving Spark headroom. |

| Routed expert gate/up, interior layers 2–45 | Q2_K, no imatrix | layer 2 Q4_K; layers 3–45 IQ2_XXS + merged activation importance matrix | Keep the aggressively quantized region broad without applying IQ2 to an under-covered layer. |

| Routed expert gate/up, protected layers | layers 0, 1, 46, 47 Q4_K | layers 0, 1, 2, 46, 47 Q4_K | Protect the first/last stages and the calibration-identified layer-2 tail. |

| Routed expert down, main 512 columns | Q4_K | Q4_K | Preserve the main expert output path. |

| Routed expert down, 128-column tail | Q4_0 | Q4_0 | Separate lower-cost tail enabled by a lossless semantic tensor split. |

| MTP routed experts and most always-active matrices | primarily Q8_0 | primarily Q8_0 | Keep speculative decoding and continuously active paths at high precision. |

| Hyper-connection, non-quantizable convolution/vision tensors | BF16 | BF16 | Preserve sensitive or shape-incompatible paths. |

| Norms, gates, recurrent/control state | F32 where required | F32 where required | Preserve numerical control behavior. |

| Integer PLE controls | I64 | I64 | Exact integer representation. |

Each metadata template contains 1,756 tensors. Relative to E5, the Q2_K track uses Q2_K instead of IQ2_XXS for 86 tensors in layers 3–45, while E5 additionally raises the two layer-2 gate/up tensors from Q2_K to Q4_K. The net Q2_K payload is 4,679,270,400 bytes / 4.36 GiB larger.

DGX Spark memory target

The provisional 256K-context projection uses:

• approximately 91.74 GiB for the published Q2_K model artifact, or 87.40 GiB for the projected calibrated E5 artifact;
• approximately 6.80 GiB for QSA KV/index state, GDN recurrent/conv state, and a conservative MTP transient allowance;
• 11 GiB reserved for runtime workspace, server/session state, and vision transients;
• an additional 8 GiB MemAvailable safety reserve.

Under an assumed 121.63 GiB usable unified-memory budget, that leaves about 4.10 GiB projected headroom for Q2_K and 8.43 GiB for calibrated E5. Raising a driver wired-memory limit does not create physical memory, so only measured residency and successful generation on one DGX Spark can pass the final memory gate.

Required validation gates

Publishing a structurally valid artifact is not the same as declaring it runtime-ready. Current gate status is:

1. Completed: generate the complete no-imatrix Q2_K GGUF and verify metadata, tensor shapes, offsets, types, local SHA-256 checksums, and remote LFS objects;
2. Completed: independently collect per-layer/per-expert routed activation importance using the official Transformers BF16 implementation, expand under-covered data, record corpus/token hashes, and adapt the precision map to the measured coverage;
3. Paused after verified template: generate and verify the complete calibrated E5 GGUF; the incomplete local shard was never uploaded;
4. compare BF16, Q8_0, and mixed-quant outputs on H200, including PLE reset behavior, GDN, QSA, four-branch gated residuals, vision, and MTP;
5. run the ds4 C/CUDA serving path and API smoke tests;
6. measure one-DGX-Spark startup, physical memory, long-context cache growth, and generation quality before declaring the release ready.

Runtime implementation references

The ds4 implementation will cross-check the official/reference serving paths where useful, especially hybrid GDN/QSA scheduling, four-branch gated residual mixing, PLE lookup/prefetch, fused MoE routing, and MTP:

• SGLang Qwen3.8-Flash-Next cookbook
• vLLM Qwen3.8-Flash-Next recipe
• TokenSpeed Qwen3.8-Flash-Next recipe
• Qwen3.8-Flash-Next official repository

These references describe BF16/FP8 serving engines, not this GGUF format. Their architecture and scheduling implementations are evidence for ds4 design decisions; they do not establish binary compatibility with this release.

Public mixed-quant references

The working method is informed by the publicly accessible releases in the DS4-Mixed-Quant-for-Spark collection:

• Baekpica/dots3-note-prev-Mixed-Quant-GGUF
• Baekpica/Motif-3-Mixed-Quant-GGUF
• Baekpica/Solar-Open2-250B-Mixed-Quant-GGUF
• Baekpica/K-EXAONE-236B-A23B-Mixed-Quant-GGUF

This Qwen release follows their general workflow—protect always-active, attention, shared-expert, control, and speculative paths; concentrate aggressive quantization in routed expert weights; and reserve unified-memory headroom for an actual server session. Its exact tensor types and validation gates remain model-specific.

Architecture retained

The target keeps the complete multimodal/speculative topology: 48 text layers (36 GDN + 12 QSA), 512 routed experts with top-10 routing, shared experts, four gated-residual branches, the 51.2B PLE n-gram table, the vision tower, and the one-layer MTP head.

License

Use of the 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.