Tosilos-128B GGUF
Community GGUF conversion and quantization of
nesilabs/tosilos-128b for use with
llama.cpp.
The original model is a Mistral Medium 3.5 128B cybersecurity fine-tune intended for authorized pentesting, vulnerability analysis, DFIR, detection engineering, and related cybersecurity workflows, as described by Nesilabs.
This repository contains a community-produced GGUF conversion/quantization. It is not a retrained or independently fine-tuned model. All credit for the original model and fine-tuning belongs to Nesilabs and the original model authors.
Available quantization
| File | Quantization | BPW | Approx. size |
|---|
tosilos-128b-Q3_K_M.gguf | Q3_K_M | 3.88 | ~60.6 GB / ~56.4 GiB |
Q3_K_M was selected as a practical starting point for running this dense 128B model on systems with approximately 128 GB of unified/system memory.
Conversion
Converted from the original BF16 Safetensors checkpoint using llama.cpp.
Conversion environment
- Source: nesilabs/tosilos-128b
- Source format: BF16 Safetensors
- llama.cpp build: 10398
- llama.cpp commit:
8e7f22b67
- Intermediate format: BF16 GGUF
- Final quantization: Q3_K_M
- Tokenizer compatibility: Transformers 5.15.0 was used successfully for conversion
The conversion pipeline was:
1nesilabs/tosilos-128b
2 |
3 | BF16 Safetensors
4 v
5convert_hf_to_gguf.py
6 |
7 v
8tosilos-128b-BF16.gguf
9 |
10 | llama-quantize Q3_K_M
11 v
12tosilos-128b-Q3_K_M.gguf
Reproducible conversion commands
1python3 convert_hf_to_gguf.py \
2 /path/to/tosilos-128b \
3 --outfile tosilos-128b-BF16.gguf \
4 --outtype bf16
1llama-quantize \
2 tosilos-128b-BF16.gguf \
3 tosilos-128b-Q3_K_M.gguf \
4 Q3_K_M
For this checkpoint, upgrading the conversion environment to Transformers 5.15.0 resolved tokenizer compatibility with the model's TokenizersBackend metadata.
Tested hardware
The Q3_K_M GGUF has been tested successfully with llama.cpp using Vulkan on:
- CPU/APU: AMD Ryzen AI MAX+ 395
- GPU: AMD Radeon 8060S Graphics
- Memory: 128 GB unified/system memory
- Backend: Vulkan
- Operating system: Ubuntu Linux
Test configuration
- Context:
32768
- KV cache:
Q8_0
- GPU layers:
999 / full layer offload
- Batch size:
512
- Micro-batch size:
256
- Parallel sequences:
1
Observed decode performance was approximately 3.7 tokens/second at a configured 32K context.
Performance will vary depending on hardware, memory bandwidth, context utilization, llama.cpp version, and runtime configuration.
Download
Using the Hugging Face CLI:
1hf download elsauto/tosilos-128b-GGUF \
2 tosilos-128b-Q3_K_M.gguf \
3 --local-dir .
Usage with llama.cpp
1llama-server \
2 -m ./tosilos-128b-Q3_K_M.gguf \
3 -c 32768 \
4 -ngl 999 \
5 -b 512 \
6 -ub 256 \
7 --cache-type-k q8_0 \
8 --cache-type-v q8_0
For systems with less available GPU or unified memory, reduce GPU offload and/or context size as necessary.
llama.cpp models.ini example
1[tosilos-128b]
2model = /models/tosilos-128b-Q3_K_M.gguf
3ctx-size = 32768
4
5n-gpu-layers = 999
6batch-size = 512
7ubatch-size = 256
8parallel = 1
9
10cache-type-k = q8_0
11cache-type-v = q8_0
12
13jinja = true
14flash-attn = auto
15
16temp = 0.3
17top-p = 0.9
18top-k = 20
19repeat-penalty = 1.05
Intended use
This GGUF preserves the intended use of the original Tosilos-128B model. Refer to the upstream model card for the authoritative description of intended uses, limitations, training information, and model-specific guidance:
This repository documents only the GGUF conversion, quantization, and tested llama.cpp configuration.
Attribution
This repository is a community-produced GGUF conversion of:
Please credit and refer to the original repository for information about the underlying model.
License
This GGUF is a quantized derivative of the original model. The original model's license and usage terms continue to apply.
Before using or redistributing this GGUF, review the license and terms published by Nesilabs in the original repository:
This repository does not grant additional rights beyond those provided by the original model's license.
