cyberneurova-Qwen3.8-Flash-Next — GGUF (imatrix (i1) quants)
🧬 Source model: cyberneurova/cyberneurova-Qwen3.8-Flash-Next
📦 Static quants: spiritfather/cyberneurova-Qwen3.8-Flash-Next-GGUF
🛠️ Runs on llama.cpp (master)
📊 Scored on CaliperBench
💬 Discord
[!Note] All credit for the model goes to cyberneurova, who produced cyberneurova-Qwen3.8-Flash-Next. These are quantizations only.
[!Tip] Benchmarked on CaliperBench — a creative-writing benchmark scoring prose craft, roleplay and willingness rather than general intelligence. See this model's scores: caliperbench.com
[!Important] Requantized with PLE (per-layer expert) mixed precision. The IQ2_M / IQ3_XXS / IQ3_M / IQ4_XS quants have been rebuilt with mixed quantization: trunk (attention, shared experts, embeddings) stays at Q8_0 while routed expert FFN tensors are quantized per-layer — up/gate projections at the named precision, down projections at IQ4_NL. This gives significantly better quality-per-byte than flat imatrix quants on MoE architectures. Each rung pairs with the external MTP draft model from the static repo. IQ4_XS is near-lossless (+0.28% PPL vs Q8_0).
This is an uncensored/unrestricted version of Qwen3.8-Flash-Next. Only the refusal behavior is removed; general model capabilities are preserved. IQ1/IQ2 quants are not available for this architecture (qwen4exp hyper-connection tensors are not covered by the imatrix calibration). Use the static Q2_K from the static repo as the lowest available quant.
imatrix (importance matrix) computed over bartowski calibration_datav3 on this model's own bf16. Weighted quants spend their bits where the calibration text says the model is sensitive; they are usually preferable to a static quant of the same size below Q5.
PLE mixed quants (new)
Requantized from bf16 with per-layer expert mixed precision. Trunk at Q8_0; routed expert tensors overridden per rung. All rungs use the same external MTP draft model (Q8_0) from the static repo.
| Quant | Size | Mixture (default / up / gate / down) | PLE | MTP pair | PPL (wikitext-2, ctx 512) | vs Q8_0 | KLD vs Q8_0 |
|---|---|---|---|---|---|---|---|
| IQ2_M | 100.9 GiB | Q8_0 / IQ2_S / IQ2_S / IQ4_NL | Q8_0 | ✓ | 5.1252 ± 0.032 | +5.97% | 0.150323 ± 0.000859 |
| IQ3_XXS | 105.6 GiB | Q8_0 / IQ3_XXS / IQ3_XXS / IQ4_NL | Q8_0 | ✓ | 4.9843 ± 0.031 | +3.05% | 0.093681 ± 0.000580 |
| IQ3_M | 109.1 GiB | Q8_0 / IQ3_S / IQ3_S / IQ4_NL | Q8_0 | ✓ | 4.9869 ± 0.031 | +3.11% | 0.072916 ± 0.000464 |
| IQ4_XS | 116.7 GiB | Q8_0 / IQ4_XS / IQ4_XS / IQ4_NL | Q8_0 | ✓ | 4.8502 ± 0.030 | +0.28% | 0.039888 ± 0.000269 |
Q8_0 baseline PPL: 4.8366 ± 0.030. Trunk (non-expert Q8_0): 55.8 GiB constant across all rungs.
Flat imatrix quants
| Quant | Size | Notes |
|---|---|---|
| Q2_K | 80.4 GB | |
| IQ3_S | 88.9 GB | |
| Q3_K_S | 88.8 GB | |
| Q3_K_M | 94.2 GB | IQ3_S probably better |
| Q3_K_L | 96.6 GB | |
| Q4_K_S | 111.8 GB | optimal size/speed/quality |
| Q4_K_M | 119.2 GB | fast, recommended |
| Q5_K_S | 127.8 GB | |
| Q5_K_M | 134.1 GB | |
| Q6_K | 167.6 GB | practically like static Q6_K |
Loading (any current llama.cpp)
llama-server -m cyberneurova-Qwen3.8-Flash-Next-IQ4_XS.gguf -ngl 99 -fa on -c 32768 -ctk q8_0 -ctv q8_0 --jinja