Upstream metadata lists Qwen/Qwen3.5-9B-Base as the base model for silx-ai/Quasar-10B; this repository is a GGUF quantization/conversion of silx-ai/Quasar-10B.

These files were converted/quantized with my llama.cpp fork:

https://github.com/Akicou/llama.cpp

They can also be run with that fork. If upstream llama.cpp does not recognize the model architecture or metadata, use the fork above.

Available files

File	Type	Size
`Quasar-10B-f16.gguf`	F16 GGUF	17,217,163,520 bytes
`Quasar-10B-Q4_K_M.gguf`	Q4_K_M quantization	5,428,039,936 bytes
`Quasar-10B-Q4_K_S.gguf`	Q4_K_S quantization	5,153,411,328 bytes

Runtime

Example using the forked llama.cpp:

bash

git clone https://github.com/Akicou/llama.cpp
cd llama.cpp
cmake -B build
cmake --build build --config Release -j

hf download Akicou/Quasar-10B-GGUF Quasar-10B-Q4_K_M.gguf --local-dir .
./build/bin/llama-cli -m Quasar-10B-Q4_K_M.gguf -p "hi"

Adjust the binary path for your platform/build type if needed.

Important behavior note

While testing, this model did not seem to be fine-tuned for instruction following or normal conversational/chat use. It appears to behave more like a continued-pretraining checkpoint. For example, even when prompted with something simple like hi, it tended to act as if it had been asked to solve a mathematical problem using constraints.

Because of that, do not expect reliable assistant-style instruction following from these files unless you add your own prompting/evaluation setup or fine-tune the model further.

Model card metadata

Base model: silx-ai/Quasar-10B
Base-model relation: quantized GGUF conversion
Upstream base listed by silx-ai/Quasar-10B: Qwen/Qwen3.5-9B-Base
License: Apache-2.0, inherited from the upstream model metadata
Languages listed upstream: English (en) and Arabic (ar)
Pipeline: text generation

GGUF metadata reported by Hugging Face Hub

The Hub detected the following GGUF metadata from the uploaded files:

Architecture: quasar
Context length: 2,097,152
EOS token: <|endoftext|>
Chat template: present in GGUF metadata
Total file size for the detected F16 GGUF: 17,217,163,520 bytes
Repo contains F16 plus Q4_K_M and Q4_K_S GGUF files