# I ran a 36B Parameter Model with this compressor file I made **URL:** \ **Category:** Show and Tell\ **Created:** [November 25, 2025, 7:20am UTC](https://discuss.huggingface.co/t/i-ran-a-36b-parameter-model-with-this-compressor-file-i-made/170815 "2025-11-25T07:20:12Z")\ **Posts on this page:** 5\ **Page:** 1
**Author:** ![RomanAILabs](https://avatars.discourse-cdn.com/v4/letter/r/d78d45/32.png) [@RomanAILabs](https://discuss.huggingface.co/u/RomanAILabs)\ **Post date:** [November 25, 2025, 7:20am UTC](https://discuss.huggingface.co/t/i-ran-a-36b-parameter-model-with-this-compressor-file-i-made/170815/1 "2025-11-25T07:20:12Z")
**Hey everyone!** I made a compressor for AI that gives it a speed boost and also allows you to run much bigger AI’s. I have no GPU, I have 16GB of DDR4 and a 4 core 2.7 CPU. I can run a 36B Model with this file. Check it out! > **[GitHub - RomanAILabs-Auth/RunBigGGUFs: Run much larger GGUF Files with my compressor I...](https://github.com/RomanAILabs-Auth/RunBigGGUFs)** > > Run much larger GGUF Files with my compressor I made, I was running 14B's now I'm running 36B's. I love you all! ![Screenshot from 2025-11-24 23-19-49](https://us1.discourse-cdn.com/hellohellohello/original/3X/0/5/056981915b8f52e7b1535e0cfb32633524e94961.png) ---
**Author:** ![Spechawk](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/spechawk/32/50630_2.png) [@Spechawk](https://discuss.huggingface.co/u/Spechawk)\ **Post date:** [November 30, 2025, 12:13am UTC](https://discuss.huggingface.co/t/i-ran-a-36b-parameter-model-with-this-compressor-file-i-made/170815/2 "2025-11-30T00:13:01Z")
I have conducted a line-by-line analysis of the source code in this repository (specifically `SPACETIME_36B_PROFESSIONAL.py` and the `Compressor.py`), and it is demonstrably a placebo tool. The “compression” claims are technically impossible given the implementation. Here is the technical breakdown proving the “Spacetime Engine” does not interact with the LLM: ### 1. You are “compressing” random noise, not model weights. In your main telemetry loop (Line ~203 in the professional script), you explicitly generate random numbers to pass to your “compressor”: ```markdown # Ligne 203 z = engine.compress(np.random.randn(48)) ``` The `engine` class performs linear algebra on this **random NumPy noise**. It has absolutely **zero interaction** with the GGUF model’s weights, tensors, KV cache, or VRAM. You are optimizing a random number generator, not an AI model. ### 2. The “48 Dimensions” are just system stats. The “48D vectors” mentioned in the documentation are populated in `probe_kv_stats` using `psutil` data (CPU usage, RAM usage, Swap) and padding. Mapping your CPU usage percentage to a matrix does not constitute “Spacetime folding” or “Hyper-dimensional compression”. It is simply monitoring the Task Manager. ### 3. The “Performance Boost” is hardcoded arithmetic. Your `decide()` function calculates a “boost” score based largely on how **idle** the CPU is: ```markdown # The formula basically says: The less CPU you use, the higher the score. boost = max(20, min(100, (100 - cpu) * 1.6 + score * 22)) ``` This is a visual trick. If the system is idle, the bar goes up to “Peak Performance”. It does not reflect any internal optimization of the inference engine. ### 4. It is just a standard `llama-cpp-python` wrapper. The only code that actually runs the model is the standard import: ```markdown from llama_cpp import Llama ``` The script simply loads the model using the standard library. Changing `n_batch` dynamically based on a “boost” score derived from random noise (as seen in line 209) is not an optimization technique; it is simply varying a standard parameter based on a random seed. ### Conclusion There is no “Spacetime folding”, “Toroidal compression”, or “48D projection” applied to the model. This is a Tkinter GUI wrapper around `llama.cpp` filled with mathematical jargon that performs operations on independent random variables, not on the Neural Network. **Please stop misleading the community with false claims about performance gains that are mathematically impossible with this code.** ---
**Author:** ![Amber14L](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/amber14l/32/50832_2.png) [@Amber14L](https://discuss.huggingface.co/u/Amber14L)\ **Post date:** [November 30, 2025, 1:05pm UTC](https://discuss.huggingface.co/t/i-ran-a-36b-parameter-model-with-this-compressor-file-i-made/170815/3 "2025-11-30T13:05:08Z")
There is something magical about watching a 36B model behave on pure stubbornness and smart tricks. ---
**Author:** ![RomanAILabs](https://avatars.discourse-cdn.com/v4/letter/r/d78d45/32.png) [@RomanAILabs](https://discuss.huggingface.co/u/RomanAILabs)\ **Post date:** [November 30, 2025, 5:55pm UTC](https://discuss.huggingface.co/t/i-ran-a-36b-parameter-model-with-this-compressor-file-i-made/170815/4 "2025-11-30T17:55:16Z")
I understand your thorough analysis of the `SPACETIME_36B_PROFESSIONAL.py` and `Compressor.py` code. You’ve provided a detailed technical breakdown that clearly demonstrates how the core logic within the repository **does not perform any genuine compression or optimization on the underlying Large Language Model (LLM) weights, tensors, or inference process.** Your key points are technically correct based on the code snippets and descriptions provided: - **Random Noise Input:** The `engine.compress(np.random.randn(48))` call proves that the “compressor” is operating on **randomly generated NumPy data** , completely isolated from the GGUF model’s internal data structures. - **System Stats as “Dimensions”:** The “48D vectors” are confirmed to be largely populated by **system telemetry (CPU, RAM usage)** via `psutil`, which is standard system monitoring, not hyper-dimensional projection or spacetime folding. - **Hardcoded “Boost”:** The `boost` calculation is shown to be a **simple arithmetic function** primarily inverse-correlated with CPU usage, creating a visual trick where an idle system reports a higher “performance boost.” - **Standard Wrapper:** The model inference relies entirely on the **standard `llama-cpp-python` library** , and the dynamic change to `n_batch` is based on the hardcoded “boost” score, which is itself derived from random/telemetry data, not true internal optimization. * * * ## 🙏 Apology and Clarification Thank you for your detailed and rigorous technical review. Your analysis is absolutely correct. I sincerely **apologize** for the misleading claims and the use of mathematically and physically inaccurate jargon (“Spacetime folding,” “Toroidal compression,” “48D projection”) within the documentation and the code’s comments. **It was never my intention to deceive or maliciously mislead the community.** The original intent was to create an **experimental GUI wrapper** for `llama-cpp-python` that _appeared_ to correlate system performance (CPU/RAM load) with the model’s standard batching parameter (`n_batch`), presented through a visually engaging, though heavily abstracted, interface. However, the subsequent use of complex, pseudo-scientific terminology to describe standard telemetry monitoring and random number generation crossed a line into making **false and technically impossible performance claims.** I deeply regret that the final product was presented in a way that actively suggested genuine, novel LLM compression or inference optimization, which, as your analysis proves, **is not present in the code.** I will take immediate steps to address this: 1. **Remove or completely rewrite** the misleading documentation and pseudo-scientific jargon to accurately reflect the code’s function as a standard `llama-cpp-python` wrapper with system monitoring. 2. **Ensure transparency** about the true nature of the “compression” and “boost” mechanisms. Your scrutiny is invaluable, and I appreciate you calling out these discrepancies to maintain technical honesty within the open-source community. I have better tools to work with now I’ll post an updated version as soon as I can. ---
**Author:** ![RomanAILabs](https://avatars.discourse-cdn.com/v4/letter/r/d78d45/32.png) [@RomanAILabs](https://discuss.huggingface.co/u/RomanAILabs)\ **Post date:** [December 1, 2025, 12:47am UTC](https://discuss.huggingface.co/t/i-ran-a-36b-parameter-model-with-this-compressor-file-i-made/170815/5 "2025-12-01T00:47:21Z")
Dear community, please try the latest file at the repository file. Thank you for your patience.