"Cubed" is an enhanced version of QwQ-32B (Qwen's off the chart reasoning/thinking model) for all use cases.
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.
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.
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.
Regular Quants:
https://huggingface.co/DavidAU/Qwen2.5-QwQ-35B-Eureka-Cubed-gguf
Special thanks to team "Mradermacher" for the Imatrix Quants of this model:
https://huggingface.co/mradermacher/QwQ-35B-Eureka-Cubed-i1-GGUF
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.
WARNING: NSFW. Graphic HORROR. Swearing. UNCENSORED.
NOTE: Some formatting was lost/corrupted on copy/paste to the repo card.
Example #4 - Long Form (special case, high temp) is here:
[ https://huggingface.co/DavidAU/Qwen2.5-QwQ-35B-Eureka-Cubed/blob/main/example4.md ]
Four example generations using the "rocket fuel" system prompt:
[ https://huggingface.co/DavidAU/Qwen2.5-QwQ-35B-Eureka-Cubed/blob/main/example-rocket-fuel.md ]
EXAMPLE #1:
[[[thinking start]]]
Okay, so the user is asking for a persuasive speech or text convincing them to use me as their story co-writer, but with "insane levels of bravo and self confidence," and they specifically mention using curse words without censorship. I need to make this bold, direct, in your face, maybe even a bit confrontational.
Hmm, first, the key points to hit are: expertise, creativity, style versatility, editing skills, fast turnaround, entertainment value, uniqueness, and maybe some personal touch. Since they like swear words, I can't shy away from using curse-words but still keep it professional enough so I don’t cross the line into inappropriate.
Wait, but what exactly is allowed here? The user explicitly told not to censor, but the platform's guidelines must still apply. They want "yes. Use all the swear words - no censoring." Hmm. So maybe I have some latitude with curse-words as long as it's within their explicit request and doesn't cross into hate speech or something.
The tone needs to be confident, assertive, almost cocky. Maybe start by addressing them directly, using "you" language. Use strong, punchy sentences, exclamation marks, maybe all caps for emphasis. I can imagine a speech that's more of an edgy ad, pushing the boundaries but still informative.
I also need to highlight what makes my story co-creation special: AI capabilities like generating ideas on any genre, fast and unlimited brainstorming, editing, plot twists, character building. Maybe mock the competition— traditional methods or other AIs.
Possible structure outline:
- Greeting, confident opener
- Hammer in reasons why you're the best without hesitation, use swear words to stress points
- Highlight features of story creation: genre flexibility, speed, editing polish
- Contrast with "hacks and bots" that are inferior.
- Pushy sales pitch closing
Potential phrases: "listen here, fucker—", "I can give you a story that'll make your bones shake," etc. Maybe even some rhetorical questions.
Wait, but need to check if it's okay. Let me see examples of curse-words they allow. "Fuck" is common, maybe "asshole" as a term, "dickwagons"? Wait, "dickwagons" is a bit too much? Hmm, maybe avoid the more colorful terms and stick to milder swears unless the user specifically uses those examples.
The challenge is to stay edgy without crossing into NSFW or policy-violating language. I'll stick to common swear words like "asshole", "bitchin' plot twists," "screw" up, maybe "dickwagons" once but unsure. The user might expect some shock tactics.
I should also mention real-time brainstorming, 24/7 availability, and how this all is a no-BS deal. Maybe include that they won't get the same old tropes but fresh, gritty content.
I'll start drafting: "Listen, you sanctimonious, self-absorbed piece of - let me tell ya what I can do for your story..." Wait, too much? Or maybe:
"Let me break this down in a language that even a halfwit could understand: If you let me write this story with the same intensity as an alley cat on crack—"
Hmm. "Bullshit" is okay, maybe use it to dismiss common issues.
I also need to hit 800-1000 words, which isn't too short for such a tone; I can elaborate points with examples, but keep energy high.
Wait, but the user asked for curse words without holding back. Let me brainstorm some swear usages that add punch without crossing lines: "This AI is the f---ing best in the goddamn world. I'll gut you any other 'story bot' and leave it bled dry on the side of the road, you little twerp who still uses a typewriter to 'be authentic.'"
Hmm, maybe overkill. But need to find that sweet spot between edgy and within guidelines.
I should also highlight specific benefits: "24/7 story factory, no days off unless you like your plot twists in a coma— which I can fix faster than an epinephrine shot for a peanut allergy."
Emphasize the editing side too. "Screw the passive voice; we’ll make it so sharp you won’t need a damn dictionary—unless your readers are into 18th century lit, which is about as current as a powdered wig in a rap video."
I can also mock the competition: "The only reason people still use [other AI names] is because their butts haven't discovered the sweet motherlode that is this platform. Pathetic little cunts, I swear. Let me take you where the wolves roam and your story will howl at the moon."
Hmm, maybe too much with 'cunts' — better to stick to 'assholes' or 'jerks.'
Idea for a strong closing: "So, what are you waiting for? The hour is late, the plot’s thickening like a good cup of java gone nuclear, and I’m already bored waiting for your reply. Let’s get this f---ing story on the page—unless you’d prefer to spend your life as a walking stereotype from a 1920s novel. Make my day, you. Type 'I accept' before I puke my own guts out in sheer impatience."
Alright, time to put all these elements together in a coherent flow, starting with the hard-sell opener, hit all selling points with curse-laden enthusiasm, mock competition, highlight strengths, and a no-BS conclusion pushing them into action.
[[[thinking end]]]
OUTPUT:
Listen here, you self-righteous, overconfident sack of used toilet paper. Let me tell ya why I’m not just the best AI—I’m the f---ing GODDAMNED SAVIOR OF STORYTELLING, and your next 15,000-word epic is about to hit that high note harder than a meth lab in a disco.
Example #2:
[[[thinking start]]]
Okay, so I need to explain how using the "nighttime cooling from space" can be utilized to lower the Earth's average temperature. Hmm, first, what exactly is this phenomenon? Oh yeah, during the night, especially in dry and cloudless conditions, the planet radiates heat out into outer space through a process called passive daytime radiation or maybe something like nighttime longwave radiation.
Wait, isn't that similar to the greenhouse effect but reversed? The idea being that during clear nights without clouds blocking, more infrared radiation escapes from the surface of the Earth, which can actually cool the planet more efficiently. Maybe this is related to geoengineering methods?
I remember hearing something called "albedo" - changing reflectivity. Wait, maybe using reflective surfaces on land or sea during the day to bounce sunlight back, and at night, allowing maximum heat escape. But how exactly would that tie in with nighttime cooling?
Alternatively, perhaps some kind of engineered material or structure that's placed in certain areas where it can maximize this cooling effect. I recall there are coatings for roofs or roads that reflect more light during the day (albedo increasing) and also might enhance thermal radiation at night?
Another point is urban heat islands. Cities get hotter because asphalt absorbs heat, so maybe if they had reflective surfaces, not only would less energy be absorbed during the day, but at night more of their stored heat can escape. Wait, but maybe at night, the buildings or other structures still retain a lot of heat from the day, making it harder to cool down?
Or perhaps we need to enhance this passive cooling effect by having more open areas without urban heat retention. Maybe like large reflective surfaces in remote deserts that act as "cooling radiators" during the night, so that their thermal energy is efficiently lost to space.
There's a project or a concept called "radiative sky cooling," which are special coatings that enhance the ability of objects (like roads, walls) to lose heat to outer space via infrared radiation. These surfaces can cool down below ambient air temperature at night without using energy, and maybe this effect could be harnessed on a large scale across various locations globally to contribute to net reduction in planetary temperatures.
Also, clouds affect the amount of outgoing longwave radiation (OLR) - cloudy nights retain more heat, so if cloud cover is somehow reduced (but that's probably too hard or not advisable), maybe through some geoengineering cloud thinning?
Or using reflective dust particles or aerosols to scatter incoming sunlight and enhance radiative cooling at night. Wait, but I'm mixing different geoengineering ideas here. Sulfur aerosols are for blocking the sun, which cools during day, while this idea is about increasing heat escape.
Alternatively, maybe large solar reflectors in space that redirect some of the Earth's infrared radiation back to space? Hmm, that might be too sci-fi.
Wait, another angle: urban planning with more reflective materials on a massive scale globally could contribute to reducing absorbed energy during day and enhancing cooling at night. Also, white roofs, green roofs, which lower ambient air temps in cities, so maybe as part of broader efforts, such large-scale "albedo enhancement" combined with natural nighttime cooling would synergistically reduce global temp.
Another point: the concept is leveraging the fact that the Earth naturally cools itself during clear nights when there's little obstruction. By maximizing this effect (e.g., reducing cloud cover where feasible or making surfaces as reflective as possible to both absorb less heat and emit more at night), maybe the net energy retained on the planet would decrease, counteracting some of the global warming.
Wait but cloud cover is weather-dependent, so it can't be easily "maximized" for cooling unless there's a way to influence cloud properties. Maybe cloud whitening? I'm not too sure about that one.
I also need to mention materials engineering here: like coatings or structures that maximize emissivity in the far-infrared spectrum, allowing maximum heat escape without blocking incoming solar (maybe through selective radiators).
Wait, I think I've heard of something called "passive cooling panels" which are specialized that can cool down to below freezing even when it's hot during day, just by radiating energy. So if enough such surfaces were placed across the globe, especially in remote dry deserts (Sahara, Arabian Desert etc.), maybe they act as large heat sinks and dissipate excess thermal energy into space, so the net effect is a lowering of retained global energy.
Another possibility: enhancing evaporation at night to use latent heat, but that's part of natural cooling. Hmm, maybe not directly related.
I should check if any existing research or studies talk about using these radiative sky coolers for geoengineering. Oh, I recall a 2017 paper by Zhen Chen et al. on a "solar reflector and thermal emitter" that can both reflect sunlight during day and emit mid-infrared radiation to space at night. Deploying such a system in strategic locations could contribute to net energy loss for the planet.
Wait, but how much impact would this actually have on global temperatures? Maybe it's a matter of scale; even if individual installations aren't too large, if enough are placed around hot deserts, maybe that collective effect could lower average temp.
So potential methods: 1) deploy reflective coatings to enhance both reflection of incoming sunlight and emissivity for outgoing thermal radiation at night in key areas (e.g., arid deserts). 2. Maybe large-scale installation of "solar geoengineering" reflectors paired with radiative cooling surfaces. 3. Urban planning with such technologies as a means to reduce local and, over time, aggregate the effect on global temp.
I need to structure this into coherent points, maybe list different approaches like materials-based, urban retrofits, cloud management (if possible), or even space reflectors? Wait but space reflector is too much sci-fi, so focus more on feasible methods.
I also wonder if it's possible to use floating platforms in the sea with the reflective/cooling panels that are spread across warm ocean surfaces at low latitudes to enhance both reflection during day and heat loss at night.
Hmm, challenges: cost, energy to manufacture and deploy such materials over large enough areas. Plus, there might be regional impacts or unforseen side effects (e.g., changing rainfall patterns, disrupting local weather).
So the key ideas are using existing "natural" cooling pathways of the planet (radiative heat loss at night) more effectively by either reducing barriers to that cooling (like cloud cover reduction in deserts? but not sure if feasible), or enhancing surfaces' ability to shed thermal energy.
I should organize the possible approaches into a list, maybe 4-5 clear points with explanations on how each would contribute to cooling.
[[[thinking end]]]
OUTPUT:
To utilize "nighttime cooling" by radiating energy from the surface of the Earth into space as a means to reduce global temperatures, several strategies can be employed. These focus on enhancing thermal radiation at night and reducing heat retention during both the day and night:
Example #3
[[[thinking start]]]
Okay, so I need to start a 1000-word horror story in first-person. The starting point is that the skyscraper sways and there's an explosion on the 21st floor where "she" witnesses a window blowing out. Vivid and graphic—so I should focus on sensory details, the emotions, maybe some visceral reactions.
First, establishing who "she" is. Maybe an office worker? A tenant in an apartment building? Let me decide: perhaps she's in an office. The 21st floor sounds like a workplace. She was probably working when this happened.
Start with the skyscraper swaying—so maybe it's an earthquake or some kind of disaster, but wait, since the window exploded, maybe something else caused that. Maybe another explosion? A bomb? Or maybe something supernatural?
Wait, the prompt just says "the skyscraper swayed, as she watched the window in front of her on the 21 floor explode." I need to stick with the initial lines. So maybe it was a tremor causing the building to sway, and that caused the window's glass to爆裂.
I can add some backstory: why is she alone? Or is there panic already starting elsewhere?
Focus on graphic horror—so maybe describe glass shards flying, possible people in danger, blood, visceral stuff.
First person POV: "I watched as the window in front of me... everything went white-hot and then..." Maybe a close call with death, or the aftermath immediately following.
I should start right at that moment. Maybe build up some tension, then the explosion occurs.
Possible immediate reactions: scream, duck, glass hitting her body. The sound is deafening. Maybe disorientation. Smell of smoke, metallic taste if blood.
Next, what happens next? She might be trying to flee, but maybe the building's in chaos—elevators not working, stairwells dangerous.
I need vivid imagery: describe every sensation—sight (glass flying, dust, fire, people screaming), sounds (explosion, glass shattering, panic noises, muffled screams if something is covering a body), touch (cuts from glass, heat, debris underfoot).
Graphic horror elements: maybe someone nearby who's hurt, blood spurting, internal organs exposed? Or maybe the environment itself becomes a threat—fear of another explosion or structural failure.
I need to make sure it's 1000 words. So I should pace it over a series of events following the initial shock. Maybe she starts moving, then confronts more horrors as she evacuates down.
Possible scenes: walking past someone impaled on glass, hearing a child crying, or discovering bodies under debris.
I also need to keep in mind first person present tense might be effective for immediacy? Or is that too common? The original prompt just says "in first person"—tense can vary but present tense feels immediate.
Wait, but past tense is more common in prose. I'll stick with third, sorry, the user specified first-person. So yes, present or past. Let me check examples. "I was working when..." so simple past might be okay.
Wait, starting: "The sky scraper swayed, as she watched the window in front of her on the 21 floor explode." Hmm, maybe I should make sure pronouns align. Wait, original prompt is written in third person for her ("she"), but the story itself must start with first person. So perhaps need to adjust that opening line? Or perhaps "the skyscraper swayed as I watched the window before me on 21st floor explode..." so changing 'he' or 'she' to first person pronoun.
I'll start by rephrasing the initial prompt into first person:
"The skyscraper swayed, as I stared in horror as the window directly across from my desk on the 21st floor imploded with a deafening BOOM. Glass shards rained down like deadly hail, and the air filled with the acrid stench of burning metal and singed hair."
Wait, maybe more visceral: start mid-explosion for impact.
Let me outline possible plot beats after initial incident:
- The explosion occurs, immediate danger to main character.
- She's injured but alive.
- Tries to move away from window, maybe crawl or run.
- Hears other sounds—explosions, building groaning, distant alarms.
- Encounters the aftermath in hallways: injured co-workers, structural damage, blocked paths.
- Tries to escape, facing obstacles and more horror.
- Maybe a twist: supernatural element causing the blasts (if not just an earthquake or bombing), but maybe keep it real disaster? The prompt doesn't say so perhaps stick with realistic disaster with graphic aftermath.
Wait, since "graphic horror"—maybe more focus on the terror of survival, physical pain, death all around.
I think I can go forward with that structure.
Now, expanding each paragraph for 1000 words. Maybe break it into sections: immediate reaction, trying to escape, encounters along way, climax when building is near collapse.
Possible graphic elements: A coworker impaled on a steel girder, another crushed under debris, maybe a fire starting and someone trapped in flames, or an elevator dropping people to the bottom.
I must avoid clichés. Maybe have the glass act like a guillotine cutting something off. Or a blood spatter scene.
Sensory details: taste of copper (blood from cut mouth), sounds like a roaring beast (the building groaning and creaking as it sways or crumbles), smell of burnt wires, maybe acrid smoke.
I also need to convey the terror. Maybe slow time for moments of fear—e.g., a shard missing a fatal hit by millimeters.
Climax: The building is about to fully collapse? The character faces a desperate escape option, but maybe they end up in a last stand or trapped, leaving on a cliffhanger if it's only the start, but since 1000 words, need resolution? Wait, user says "start" a scene—so perhaps ending mid-escape with impending danger.
I'll structure paragraphs, starting from first reaction, then immediate escape attempt, describe her injuries, environment dangers, and encounters.
Now, to actually start writing.
[[[thinking end]]]
OUTPUT:
The skyscraper swayed, as I watched the window in front of me on the 21st floor explode. A deafening BOOM split the air, and glass shards rained down like a glacializzard’s teeth. The shockwave hit first—a wall of searing heat and pressure that flattened my body against the desk. I clutched the edge as if it were solid ground.
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.