---

base_model: Qwen/Qwen3.8-27B

license: apache-2.0

library_name: llama.cpp

pipeline_tag: image-text-to-text

tags:

• gguf
• qwen3_5
• multimodal
• quantized
• 1-bit

---

Qwen3.8-27B — 1-bit-class GGUF family

Ultra-low-bit GGUF quantizations of Qwen/Qwen3.8-27B, produced directly from a verified BF16 GGUF source with llama.cpp and an agentic-domain importance matrix.

These are post-training 1-bit-class quants, not BitNet and not a model trained with quantization-aware training. The format names describe the low-bit tensor quantizer, while llama.cpp deliberately keeps sensitive tensors at higher precision. Therefore the effective model-wide bits-per-weight are higher than the nominal label.

Variants

| Variant | Quantizer label | Nominal bpw | Direct-BF16 dry-run effective bpw | Notes |

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

| Q1_0 | Q1_0 | 1.125 | ~1.38 | Smallest and most aggressive control/release. Expect the largest quality loss. |

| Q1_0-HYBRID-IQ2 | experimental mixed Q1_0/IQ2 | 1.125 base | ~2.03 | Experimental: protects attention, SSM and FFN-down pathways with extra bits. |

| TQ1_0 | TQ1_0 | 1.69 | ~2.05 | Ternary quantization. |

| IQ1_S | IQ1_S | 1.56 | ~2.12 | Imatrix-guided IQ quant. |

| IQ1_M | IQ1_M | 1.75 | ~2.26 | Imatrix-guided, largest of this family and generally the least aggressive IQ1 choice. |

The effective values above are llama.cpp dry-run values for this exact model and toolchain, not generic marketing figures.

Provenance

• Base model: Qwen/Qwen3.8-27B
• Base revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
• Architecture in GGUF: qwen35
• License: Apache-2.0
• BF16 GGUF source: unsloth/Qwen3.8-27B-GGUF
• BF16 revision: f1bfb127c64f7072bdd2cad55f258b9c8b2910fe
• BF16 split SHA-256:

- b9966e82b7a4d87028b5eae061d578ee826305ebf8baea5bfc6e09bad0ba191f

- 92e3943c4f9bd6292a7bef82369f65fed9bfed088b9df0fb2fa2ce17c9edfa02

• Importance matrix: KikoCis/Qwen3.8-27B-GGUF, revision f5dd0ad6832693dea3a648b040e2ac53e29668fb
• Imatrix SHA-256: 20334fd9bd7beb3ecc85df231433b29a042a124a3ab7304599a6f330ba4290e9
• Multimodal projector: pinned Unsloth mmproj-BF16.gguf
• Projector SHA-256: 83ee4f4f205fa514161778c41df1ea14144faa0f713510893b63c2395f5c2d53
• Quantizer: llama.cpp build 9591, commit 62061f910
• Quantization host: existing Hetzner GETH CPU server; no dedicated GPU quantization host was required.

All final family variants are quantized directly from BF16, rather than from the temporary Q8 control conversion.

Experimental hybrid

Q1_0-HYBRID-IQ2 uses Q1_0 as the default but selectively protects Qwen3.8's recurrent/attention pathways. The recipe raises attention Q/K/V/output, linear-attention QKV/gate/output and FFN-down matrices to IQ2_XS, keeps the small SSM alpha/beta matrices at Q8_0, and keeps token embeddings at Q2_K. The exact regex override recipe is included under experimental/.

This is an experiment, not an upstream llama.cpp preset. It is included because its dry-run size is slightly below TQ1_0 while deliberately allocating additional precision to pathways likely to amplify quantization error.

Accuracy warning

These are extreme compression formats. Reasoning, factual accuracy, coding, instruction following, multilingual capability and vision quality may degrade sharply, especially with Q1_0. Use a larger quantization when correctness matters. Do not interpret a successful load test as evidence that the model retains full BF16 capability.

Qwen3.8's current llama.cpp CPU SSM path can also make prompt processing unusually slow. Runtime latency and artifact integrity are therefore recorded separately in the validation reports.

Vision

The shared mmproj-BF16.gguf is provided for multimodal use with a recent llama.cpp build supporting qwen35 multimodal models.

Verification

Each model upload is gated on GGUF magic, a complete GGUF structural/tensor parse, SHA-256 and llama.cpp GGUF hashing, architecture metadata, and remote Hugging Face byte-size plus range/magic verification. Runtime text/vision probes are recorded separately because CPU SSM execution speed can be extremely slow on this architecture.

Per-variant hashes, byte sizes and source details are collected in FAMILY_MANIFEST.json and per-variant provenance/validation files.