Instructions to use Harsha001/SenseNova-Loop-9B-ROCmFPX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use Harsha001/SenseNova-Loop-9B-ROCmFPX with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP # Run inference directly in the terminal: llama cli -hf Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP # Run inference directly in the terminal: llama cli -hf Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP # Run inference directly in the terminal: ./build/bin/llama cli -hf Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
Use Docker
docker model run hf.co/Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
- Llama new
- LM Studio
- Jan
- vLLM
How to use Harsha001/SenseNova-Loop-9B-ROCmFPX with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Harsha001/SenseNova-Loop-9B-ROCmFPX" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Harsha001/SenseNova-Loop-9B-ROCmFPX", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
- Ollama
How to use Harsha001/SenseNova-Loop-9B-ROCmFPX with Ollama:
ollama run hf.co/Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
- Unsloth Desktop
- Pi
How to use Harsha001/SenseNova-Loop-9B-ROCmFPX with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:9931/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use Harsha001/SenseNova-Loop-9B-ROCmFPX with Docker Model Runner:
docker model run hf.co/Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
- Lemonade
How to use Harsha001/SenseNova-Loop-9B-ROCmFPX with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
Run and chat with the model
lemonade run user.SenseNova-Loop-9B-ROCmFPX-Q4_0_ROCMFP
List all available models
lemonade list
- Hermes Agent
How to use Harsha001/SenseNova-Loop-9B-ROCmFPX with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:9931/v1 hermes config set model.default Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use Harsha001/SenseNova-Loop-9B-ROCmFPX with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:9931/v1 \ --custom-model-id "Harsha001/SenseNova-Loop-9B-ROCmFPX:Q4_0_ROCMFP" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
Download README.md from Harsha001/SenseNova-Loop-9B-ROCmFPX: direct link, hf CLI and curl.
- Browser
- Download file 4.57 kB
-
https://huggingface.co/Harsha001/SenseNova-Loop-9B-ROCmFPX/resolve/main/README.md
- Command line
-
hf download hf://Harsha001/SenseNova-Loop-9B-ROCmFPX/README.md
-
curl -L -o README.md https://huggingface.co/Harsha001/SenseNova-Loop-9B-ROCmFPX/resolve/main/README.md
license: apache-2.0
base_model: sensenova/SenseNova-Loop-9B
base_model_relation: quantized
pipeline_tag: image-text-to-text
library_name: gguf
language:
- en
- zh
tags:
- gguf
- rocmfpx
- amd
- vulkan
- strix-halo
- quantized
- looped-transformer
- mhc
- vision-language
- reasoning
SenseNova-Loop-9B, ROCmFPX GGUF
Quantized files of SenseNova-Loop-9B, a 9B vision-language model from SenseNova. It runs 24 of its 32 decoder layers twice (56 layer passes per token) and mixes two residual streams inside the looped part (mHC). It takes text and images and writes text.
These files need HarshaVardhan3002/ROCmFPX, built from 18 November 2026 or later. The
sensenova_looparchitecture is new. Stock llama.cpp, LM Studio, Ollama and older ROCmFPX builds will not load any file here, the 6-bit one included.Not tested on AMD hardware yet. Please report speed and quality in the Community tab.
Not an official release. All credit for the model goes to the SenseNova team.
Files
| File | Size | KL to BF16 | Same top token | |
|---|---|---|---|---|
SenseNova-Loop-9B-Q8_0_ROCMFPX.gguf |
9.3 GB | 0.0012 | 98.7% | 8-bit, start here |
SenseNova-Loop-9B-Q6_K_ROCMFPX.gguf |
7.4 GB | 0.013 | 95.5% | 6-bit ROCmFPX (Q6_0_ROCMFPX), added 20 November |
SenseNova-Loop-9B-Q4_0_ROCMFP4_STRIX_O8.gguf |
6.6 GB | not measured yet | 4-bit for AMD Strix Halo: ROCmFP4-FAST, attention and output 8-bit, added 20 November | |
mmproj-SenseNova-Loop-9B-F16.gguf |
0.9 GB | vision encoder, needed for images |
For comparison, measured the same way:
| Not published | Size | KL to BF16 | Same top token |
|---|---|---|---|
| Stock Q6_K | 7.4 GB | 0.012 | 95.9% |
| Stock Q8_0 | 9.5 GB | 0.0023 | 98.5% |
The ROCmFPX 8-bit file is smaller than a stock Q8_0 and about half as far from BF16. At 6-bit, the stock Q6_K is slightly closer (0.012 against 0.013); the ROCmFPX 6-bit runs on the fork's AMD kernels. Given how sensitive this model is, the 4-bit Strix file is for speed and memory on Strix Halo, not for the closest match; it passed a short answer test ("The capital of France is Paris."), and its numbers will be added.
This model reacts more strongly to quantization than a plain 9B model with the same layer layout: the same recipes on a Qwen3.5-9B fine-tune drifted about seven to ten times less. A likely reason is the loop, where rounding errors in the shared layers pass through them twice. If you need the closest match to the original, use the 8-bit file.
How this was measured: llama.cpp llama-perplexity --kl-divergence against a BF16 GGUF converted from SenseNova's
weights (perplexity 14.46), 40 chunks of 512 tokens of English Wikipedia text (wikitext-2), on an NVIDIA GB10 GPU after
expanding each file to F16. KL is how far the file's next-token probabilities drift from BF16, averaged over every
token; lower is better. This is general English text, not reasoning, coding or agent tasks, so it says nothing direct
about the benchmarks on the original card.
Port check: the new architecture in the fork was compared with SenseNova's own PyTorch code on the full model. On a 37-token chat prompt, the F32 GGUF gives the same next-token probabilities as PyTorch in FP32 (cosine 0.9999998, largest KL 0.0000012, same top token at every position). A small random model with the same layout also matched token by token with the cache.
Short checks on the published files, on CPU with greedy decoding:
- Text: both answered "What is the capital of France?" with "The capital of France is Paris."
- Vision: the 8-bit file correctly described a test image as a red circle in the upper left and a blue square in the lower right on a white background.
Notes:
- The multi-token prediction (MTP) head is not included.
- The model card's chat template, with its thinking block, is in the files. To answer without thinking, start the
answer with an empty
<think>\n\n</think>\n\nblock. - The loop and the stream mixing turn each token into many small GPU operations, so generation is slower than a plain 9B model of the same size.
Credits and license
- Original model: sensenova/SenseNova-Loop-9B, Apache-2.0, code at OpenSenseNova/SenseNova-Loop.
- The fork's support reuses llama.cpp's Qwen3.5 code, since the decoder uses the same hybrid layer layout.
Apache-2.0. See LICENSE.