wrldsuksgo2mars/Qwen3.8-Flash-Next-EXL3-K4.25-v1

Qwen3.8-Flash-Next EXL3 K4.25 v1

Calibrated mixed K4/K5 EXL3 routed experts, averaging 4.25 bits per expert weight, including all 48 target layers and the native MTP layer. Per layer, K5 upgrades cover 72 gate, 120 up and 192 down projections; remaining expert projections use K4. This average describes routed expert storage, not the whole checkpoint.

Original non-expert tensors are retained, including BF16 PLE n-gram tables. The 22 safetensors shards contain 302,616 tensors and 179,366,664,184 bytes of tensor payload (179,406,710,432 bytes including shard headers). The original license, tokenizer, chat template and processor assets are included.

Compatibility

Produced with the GPTQModel fork at 339cccf0c4b62440b572440a7d06608ed9e4a165. Native execution is being developed in QFlashRT for RTX PRO 6000 Blackwell and DGX Spark. The runtime exports separate native artifacts from this checkpoint and reads PLE tables from disk. This is not a claim of plug-and-play support in stock Transformers or every EXL3 loader.

Expert activations remain 16-bit. QFlashRT's balanced/accurate profiles select FP8/BF16 KV respectively; KV storage is not part of this quantized checkpoint. Vision weights and original processor assets are retained, but end-to-end vision and complete serving qualification remain in progress.

Validation

An independent byte audit checked all 301,056 packed expert tensors against 49 compact runtime expert artifacts and all 1,560 non-expert tensors against the authenticated original source. Configuration, tokenizer/processor assets and all 75,264 packed projection descriptions were audited separately.

An exploratory held-out evaluation covers 89 records and 67,477 next-token predictions, using original framework block math with the GPTQModel EXL3 projection backend. These runs evaluate only the 48 target layers, with MTP disabled. The EXL3 MTP layer is included in this release but needs separate acceptance, quality and speed evaluation. Lower perplexity is better on this fixed test:

Device	Original BF16 reference	EXL3 framework reference
RTX PRO 6000 Blackwell	3.87319	3.87893
DGX Spark / GB10	3.89576	3.85601

These results include backend numerical effects, are hardware-specific, and are not a general model ranking or task-quality certification. Coding, reasoning, long-context, tools, vision and MTP quality qualification continues.

Provenance

• Original model revision: de4b8e4d43b917e7706784d8bb445c9af86a3540.
• Source identity: 1173fcfcb79c115f9a9339219fc6386bdda636f49fa42cb52e86d2b36045679f.
• Quantization run: d8413370b3e143a6925d73f5ff441eee9d832506ba1ff66f360b0362219f7bae.
• Safetensors index SHA256: 4be63d793de784e4e89b1a1aba3809109b3c88c87f58f0b79b88251000c5c148.
• Calibration and runtime details: quantization documentation.

The source Qwen Community License 1.0 applies; see the included LICENSE.