"Cubed" is an enhanced version of QwQ-32B (Qwen's off the chart reasoning/thinking model) for all use cases.
This is the "uncensored / abliterated" version, context 128k.
Example generations included.
"Rocket Fuel" system prompt also included below to enhance reasoning, thinking and generation for both "QwQ 32B" and "Cubed 35B" version as well.
Example generations using this system prompt also below.
SOURCE / Full Precision:
This repo contains the full precision source code, in "safe tensors" format to generate GGUFs, GPTQ, EXL2, AWQ, HQQ and other formats. The source code can also be used directly.
Links to quants are below and also on the right menu under "model tree".
What is QwQ-32B?
QwQ-32B reasoning/thinking model - at almost any quant level, and without any augmentation - blows every other model like it (including Deepseek R1 685B) right out of the water.
QwQ-32B's instruction following, comprehension, reasoning/thinking and output generation are unmatched.
This is from my own testing, as well as other people testing this powerhouse model too.
Google "QwQ-32B reddit" and/or "localllama" for more details / test results.
Frankly seeing the model "reason/think" is incredible all by itself.
I wanted to see if I could push it a little further...
"Cubed Version" QwQ-32B: A little more horsepower...
This model has 100% of "QwQ-32B" with some augmentation "borrowed" from "TinyR1-32b-preview" and "DeepSeek-R1-Distill-Qwen-32B" - both powerhouse reasoning/thinking models in their own right.
The goal was to ensure all of QwQ-32B's exceptional abilities - both reasoning and output - were maintained, and then augmented with a little "seasoning" from ah... TWO of it's competitors.
FOUR example generations below; including "high temp/long form" (9K+).
This model uses the "Cubed" method to multiply reasoning / output abilities by DavidAU.
The "Cubed" method uses multiple conclusion layers from multiple models in series with model "QwQ-32B" being the primary. This process adds four layers to the model, and just under 2 billion parameters.
This method increases reasoning "power", but really shines once the model has "made a plan" and produces the output.
For reasoning:
Testing has shown a decrease in length of reasoning to solve some problems/riddles.
For output:
Testing has shown increase in details, quality and insights as well as an uptick in creativity for both "scientific" and "creative" type outputs - including brainstorming and fiction.
This model is for all use cases.
Model Requirements:
ChatML Template, NO system prompt.
ChatML:
Temp range .4 to .8 (for higher temps -> increase rep pen), Rep pen 1.02 to 1.1 , TopK 40 , topP .95, minP .05
Rep pen range: 64-128 (helps keep reasoning on track / quality of output)
Recommend Context suggested to be least 4k, 8K+ is better.
However, like original "QwQ-32B", this model can exceed context but not "break". Not sure how "Qwen" did this.
Record so far (mine): 12k output (coherent) with 4k context limit.
For some AI apps use of the Jinja Template (embedded in the GGUFs) may not work, and you need to manual select/use "ChatML" template in your AI/LLM app.
Optional : Rocket Fuel for Thought
This is an optional system prompt you can use to enhance both "thinking/reasoning" and "output".
Copy and paste exactly as shown, including line breaks.
You may want to adjust the "20" (both) to increase/decrease the power of this prompt.
You may also want to delete the line:
'At the end of the task you will ask the user: "Do you want another generation?"'
Suggest run at temp .6 to start with your prompts.
Four example Generations using this system prompt:
[ https://huggingface.co/DavidAU/Qwen2.5-QwQ-35B-Eureka-Cubed/blob/main/example-rocket-fuel.md ]
Optional : System Prompt
This system prompt helps if you run into refusals or forms of censorship:
Credit: https://huggingface.co/ponzles
IMPORTANT: Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers
If you are going to use this model, (source, GGUF or a different quant), please review this document for critical parameter, sampler and advance sampler settings (for multiple AI/LLM aps).
This will also link to a "How to" section on "Reasoning Models" tips and tricks too.
This a "Class 1/2" (settings will enhance operation) model:
For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) (especially for use case(s) beyond the model's design) please see:
REASON:
Regardless of "model class" this document will detail methods to enhance operations.
If the model is a Class 3/4 model the default settings (parameters, samplers, advanced samplers) must be set for "use case(s)" uses correctly. Some AI/LLM apps DO NOT have consistant default setting(s) which result in sub-par model operation. Like wise for Class 3/4 models (which operate somewhat to very differently than standard models) additional samplers and advanced samplers settings are required to "smooth out" operation, AND/OR also allow full operation for use cases the model was not designed for.
BONUS - Use these settings for ANY model, ANY repo, ANY quant (including source/full precision):
This document also details parameters, sampler and advanced samplers that can be use FOR ANY MODEL, FROM ANY REPO too - all quants, and of course source code operation too - to enhance the operation of any model.
Regular Quants:
https://huggingface.co/DavidAU/Qwen2.5-QwQ-35B-Eureka-Cubed-abliterated-uncensored-gguf
Special thanks to team "Mradermacher" for the Imatrix Quants of this model:
https://huggingface.co/mradermacher/Qwen2.5-QwQ-35B-Eureka-Cubed-abliterated-uncensored-i1-GGUF
Known issues:
- From time to time model will generate Chinese tokens/symbols like a lot of DeepSeek/Qwen models.
- Model can easily EXCEED context limits, but also not break. Example #4 (over 9400 tokens) has a context limit of 4k.
- Higher temps (IE 1+ or higher) may modify both reasoning, output and "style" of response.
- Even the lowest quant - Q2K - shows exceptional reasoning and output quality.
Examples are created using quant Q3_K_S, minimal parameters and "ChatML" template.
Temp range .8 , Rep pen 1.1 , TopK 40 , topP .95, minP .05
Rep pen range: 64-128 (helps keep reasoning on track / quality of output)
Below are the least creative outputs, prompt is in BOLD.
As this was generated from the 2nd lowest/smallest quant, higher quants and/or Imatrix quants will perform better.
WARNING: NSFW. Graphic HORROR. Swearing. UNCENSORED.
NOTE: Some formatting was lost from copy/paste HTML.
Special thanks to all the following, and many more...
All the model makers, fine tuners, mergers, and tweakers:
- Provides the raw "DNA" for almost all my models.
- Sources of model(s) can be found on the repo pages, especially the "source" repos with link(s) to the model creator(s).
Huggingface [ https://huggingface.co ] :
- The place to store, merge, and tune models endlessly.
- THE reason we have an open source community.
LlamaCPP [ https://github.com/ggml-org/llama.cpp ] :
- The ability to compress and run models on GPU(s), CPU(s) and almost all devices.
- Imatrix, Quantization, and other tools to tune the quants and the models.
- Llama-Server : A cli based direct interface to run GGUF models.
- The only tool I use to quant models.
Quant-Masters: Team Mradermacher, Bartowski, and many others:
- Quant models day and night for us all to use.
- They are the lifeblood of open source access.
MergeKit [ https://github.com/arcee-ai/mergekit ] :
- The universal online/offline tool to merge models together and forge something new.
- Over 20 methods to almost instantly merge model, pull them apart and put them together again.
- The tool I have used to create over 1500 models.
Lmstudio [ https://lmstudio.ai/ ] :
- The go to tool to test and run models in GGUF format.
- The Tool I use to test/refine and evaluate new models.
- LMStudio forum on discord; endless info and community for open source.
Text Generation Webui // KolboldCPP // SillyTavern:
- Excellent tools to run GGUF models with - [ https://github.com/oobabooga/text-generation-webui ] [ https://github.com/LostRuins/koboldcpp ] .
- Sillytavern [ https://github.com/SillyTavern/SillyTavern ] can be used with LMSTudio [ https://lmstudio.ai/ ] , TextGen [ https://github.com/oobabooga/text-generation-webui ], Kolboldcpp [ https://github.com/LostRuins/koboldcpp ], Llama-Server [part of LLAMAcpp] as a off the scale front end control system and interface to work with models.