language:
- en
- zh
license: apache-2.0
pipeline_tag: image-text-to-text
tags: - cyinnove
- gguf
- qwen3
- conversational
- reasoning
- vision
base_model: - DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF
- DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU
- Qwen/Qwen3.8-27B
CyInnove Qwen3.8 27B
Maintained by CyInnove. Cyber innovation to secure assets and applications.
This repository is the CyInnove inference copy of a public Qwen3.8 27B GGUF release. The weights in this repo are unchanged from that clone. CyInnove has not fine-tuned them yet. This card is the baseline those later runs will start from.
Credits
Cloned from DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF.
That release quantizes DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU, a fine-tune of Qwen/Qwen3.8-27B (about 27.8B parameters, qwen3_5, AutoModelForMultimodalLM).
Upstream training, quantization, and the benchmark write-up belong to DavidAU and the collaborators named on that model card, including Nightmedia, TeichAI, armand0e, trohrbaugh, and nbeerbower. Upstream datasets referenced there include DavidAU/Polar-STRICT-Datasets and DavidAU/F451-STRICT-Datasets. This repository does not repeat those benchmark claims.
License: Apache-2.0, inherited from the upstream release. Redistribution must keep this credit and the license.
What changed in this copy
- Publisher and model card are CyInnove.
- Weight files are the cloned GGUF set, byte-for-byte. Filenames are still the upstream names.
- No CyInnove dataset has been applied.
Files
Text quants use the prefix Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX.
| Family | Quant | Size |
|---|---|---|
| Standard | IQ2_M | 11.7 GB |
| Standard | IQ3_M | 14.1 GB |
| Standard | IQ4_XS | 16.6 GB |
| Standard | Q4_K_S | 17.1 GB |
| Standard | Q4_K_M | 18.0 GB |
| Standard | Q5_K_S | 20.2 GB |
| Standard | Q5_K_M | 20.7 GB |
| Standard | Q6_K | 23.6 GB |
| Standard | Q8_0 | 29.8 GB |
| MTP | IQ2_M | 12.1 GB |
| MTP | IQ3_M | 14.5 GB |
| MTP | IQ4_XS | 17.0 GB |
| MTP | Q4_K_S | 17.5 GB |
| MTP | Q4_K_M | 18.5 GB |
| MTP | Q5_K_S | 20.6 GB |
| MTP | Q5_K_M | 21.2 GB |
| MTP | Q6_K | 24.0 GB |
| MTP | Q8_0 | 30.2 GB |
| Low MTP | IQ4_XS | 15.3 GB |
| Low MTP | Q6_K | 22.4 GB |
Vision projectors, loaded beside a text quant:
| File | Size |
|---|---|
mmproj-F16.gguf |
928 MB |
mmproj-BF16.gguf |
931 MB |
mmproj-F32.gguf |
1.8 GB |
Use one text quant. Q5_K_M is the default local check. Q4_K_M is the practical floor when tool calling matters. Q6_K and Q8_0 are the higher-fidelity checks. MTP files are the upstream multi-token-prediction quants. Low MTP is the shorter-thinking pair from the same release.
Run
llama-server \
-m Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-Q5_K_M.gguf \
--mmproj mmproj-F16.gguf \
-c 8192
Drop --mmproj for text-only use. The GGUF carries the upstream chat template. llama.cpp applies it.
Fine-tune, later
These GGUF files are for inference. They are not the training checkpoint.
When CyInnove fine-tunes this line:
- Train on the safetensors checkpoint, DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU, not on a quant in this repo.
- Use supervised fine-tuning with QLoRA (4-bit base, trainable adapters only) and TRL
SFTTrainer. - Accept data as JSONL conversations:
{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
- Hold out 20 percent for eval. Publish the adapter as its own repo under
zomasec, then merge and build a new GGUF. - Leave the files in this repo as the frozen baseline until that GGUF exists. The CyInnove fine-tune will be labeled as such on the card. Until then, every file here is the clone.