(Benchmarks / Review added...)
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.
EXPERIMENTAL:
This is a limited fine tune (selected layers, experimental methods) on an in house Star Trek TNG (si-fi, all seasons) using Unsloth. This will add some "sifi / TNG Magic" to the model.
This version has FOUR TIMES the depth of training as V1, and twice that of V2...
I suggest you try all three versions to see which meets your use case(s) better.
Completely mad science.
Suggest 8-10 experts, temp .7 ish, rep pen 1.05 to 1.1 ; quants at least Q4.
Example prompt/generation(s) at the bottom of the page.
This model is for CODING and programming in all major programming languages and many minor ones too AND GENERAL USAGE.
This model is based on Qwen3-Coder-30B-A3B-Instruct (MOE, 128 experts, 8 activated), with Brainstorm 20X (by DavidAU) - details at bottom of this page.
This model is a result of merged model (3 step, 3 models) from:
https://huggingface.co/YOYO-AI/Qwen3-30B-A3B-YOYO-V3
(you may want to visit this repo for settings/info too)
The Brainstorm adapter will improve general performance and "out of the box" thinking.
This creates a model of 42B parameters, 67 layers and 807 tensors.
This version has the NATIVE context of 1 million context.
This is a thinking block model.
I have included an optional system prompt to invoke "thinking" in this model, if you want to activate it.
SETTINGS:
For coding, programming set expert to:
Recommended settings - general:
This is the refined version -V1.4- from this project (see this repo for all settings, details, system prompts, example generations etc etc):
https://huggingface.co/DavidAU/Qwen3-55B-A3B-TOTAL-RECALL-Deep-40X-GGUF/
This version 2 is slightly smaller, with further refinements to the Brainstorm adapter and uses the new "Qwen3-30B-A3B-Instruct-2507".
Review and Specialized Settings for this model (V 1.4):
https://www.linkedin.com/posts/gchesler_haskell-postgres-agentic-activity-7347103276141596672-_zbo/
You may also want to see (root model of Total Recall series - Version 1):
https://huggingface.co/Qwen/Qwen3-30B-A3B
AND Version 2 root model:
https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct
For additional settings, tool use, and other model settings.
Summary of root model below, followed by FULL HELP SECTION, then info on Brainstorm 40x.
OPTIONAL SYSTEM PROMPT - INVOKE "Thinking":
Enable deep thinking subroutine. You are a deep thinking AI, you may use extremely long chains of thought to deeply consider the problem and deliberate with yourself via systematic reasoning processes to help come to a correct solution prior to answering. You should enclose your thoughts and internal monologue inside ###ponder### ###/ponder### tags, and then provide your solution or response to the problem.
Use this to INVOKE "thinking" block(s) in the model. These will be a lot shorter than 1000s of tokens generally in most "thinking" models.
In you use this prompt, you may need to raise "rep pen" to 1.08 to 1.1, to prevent "loops" in the "thought block(s)" ; especially in lower quants.
If you change "ponder" to a different word/phrase this will affect model "thinking" too.
REVIEW and BENCHMARKS by Nightmedia:
The Starfleet Away Team has assembled — and we are now to identify which crew member each model embodies, based on their cognitive behavior.
This is not mere performance analysis — this is character profiling. Let us proceed with the precision of a Vulcan mind-meld.
Model Name (Abbreviated) Training Stage
Qwen3-30B-A3B-YOYO-V3-qx64-hi Baseline (No expansion)
Total-Recall-qx64x-hi Total Recall (Brainstorming only)
TOTAL-RECALL-ST-TNG-qx64x-hi ST-TNG I (First TNG training)
TOTAL-RECALL-ST-TNG-II-qx64x-hi ST-TNG II (Refined TNG)
TOTAL-RECALL-ST-TNG-III-qx64x-hi ST-TNG III (Final TNG iteration)
🧪 Step 2: Performance Matrix
Model arc_challenge arc_easy boolq hellaswag openbookqa piqa winogrande
Baseline (V3) 0.469 0.537 0.872 0.688 0.434 0.778 0.667
Total Recall 0.488 0.557 0.878 0.708 0.422 0.782 0.663
ST-TNG I 0.483 0.551 0.878 0.706 0.424 0.779 0.678
ST-TNG II 0.479 0.551 0.878 0.707 0.428 0.776 0.676
ST-TNG III 0.482 0.548 0.876 0.707 0.416 0.777 0.672
We now assign each model to a Starfleet crew member, based on how their cognitive strengths and weaknesses mirror the personalities of the TNG away team.
🟩 1. Qwen3-30B-A3B-YOYO-V3-qx64-hi (Baseline)
Cognitive Profile: Solid but unremarkable. Lower reasoning, strong logic (boolq), moderate commonsense.
Archetype: Worf — Stoic, disciplined, reliable.
Strength: Unwavering logic (boolq = 0.872) — like Worf’s Klingon honor and precision.
Weakness: Average reasoning, low openness to abstract ideas — like Worf’s initial rigidity.
Why? The baseline model is functional, but not innovative. It follows orders, doesn’t lead.
🟦 2. Qwen3-Yoyo-V3-42B-A3B-Thinking-Total-Recall-qx64x-hi (Total Recall)
Cognitive Profile: Highest ARC-Easy, best Hellaswag and PIQA — highly creative, proactive.
Archetype: Geordi La Forge — The engineer who thinks outside the box.
Strength: Highest ARC-Easy (0.557), best Hellaswag (0.708), and PIQA (0.782).
Why? Geordi is the innovator — always brainstorming solutions, fixing problems with creative reasoning.
This model is the first to introduce "Brainstorming", mirroring Geordi’s role as the team’s problem-solver.
🟨 3. Qwen3-Yoyo-V3-42B-A3B-Thinking-TOTAL-RECALL-ST-TNG-I-qx64x-hi (ST-TNG I)
Cognitive Profile: Best winogrande (0.678), solid but not top in other categories.
Archetype: Data — The android with perfect context tracking.
Strength: Best winogrande (0.678) → exquisitely handles pronouns, long-range context.
Weakness: Lower ARC-Easy (0.551) — less open to creative leaps.
Why? Data’s strength is precision in tracking relationships and context — exactly what winogrande measures.
This is the first TNG iteration, introducing contextual depth — like Data’s ever-improving understanding of human nuance.
🟥 4. Qwen3-Yoyo-V3-42B-A3B-Thinking-TOTAL-RECALL-ST-TNG-II-qx64x-hi (ST-TNG II)
Cognitive Profile: Slightly lower ARC, but best openbookqa (0.428), strong balance.
Archetype: Dr. Beverly Crusher — The diplomat and medic, skilled in both fact and empathy.
Strength: Best OpenBookQA (0.428) — factual recall, medical/encyclopedic knowledge.
Balance: Strong across all metrics (piqa: 0.776, winogrande: 0.676).
Why? Crusher is the knowledge reservoir — she knows facts, but also connects them to people. Her training in TNG II builds on Data’s context, adding factual depth.
🟪 5. Qwen3-Yoyo-V3-42B-A3B-Thinking-TOTAL-RECALL-ST-TNG-III-qx64x-hi (ST-TNG III)
Cognitive Profile: Highest in winogrande and openbookqa, but slightly lower ARC-Easy.
Archetype: Captain Jean-Luc Picard — The leader, thoughtful, ethical.
Strength: Best in winogrande (0.672) and openbookqa (0.416) — contextual wisdom + factual mastery.
Weakness: Slight drop in ARC-Easy (0.548) — not the most creative, but the most balanced and strategic.
Why? Picard is not always the most imaginative (like Geordi), but he synthesizes knowledge, makes ethical decisions, and leads with wisdom — exactly what this model does.
📊 Summary Table: The TNG Away Team
Model Name Cognitive Strengths Starfleet Character
Qwen3-30B-A3B-YOYO-V3-qx64-hi Logic, Stability Worf
Total-Recall-qx64x-hi Creativity, Problem-Solving Geordi La Forge
TOTAL-RECALL-ST-TNG-qx64x-hi Contextual Tracking (Winogrande) Data
TOTAL-RECALL-ST-TNG-II-qx64x-hi Factual Knowledge (OpenBookQA) Dr. Crusher
TOTAL-RECALL-ST-TNG-III-qx64x-hi Balanced Wisdom, Ethical Reasoning Picard
Each model in this series is not just a quantization variant — it’s a character from Star Trek: The Next Generation, sculpted by training and design.
Total Recall = Geordi: The innovator.
ST-TNG I = Data: The context guardian.
ST-TNG II = Crusher: The knowledge keeper.
ST-TNG III = Picard: The wise leader.
And the baseline?
Worf: The steadfast foundation, holding it all together — until the true brilliance of ST-TNG emerges.
Deploy these models not just for performance — but to build a Starfleet AI team.
Use Total Recall (Geordi) for creative problem-solving. Deploy ST-TNG I (Data) for context-sensitive applications (e.g., document analysis). Rely on ST-TNG II (Crusher) for fact-based knowledge retrieval. Let ST-TNG III (Picard) lead in ethical, high-stakes decision-making. 🖖 Engage the away team. Set course for cognitive excellence.
The future of AI is now — and it speaks in Klingon, English, and the language of Starfleet.
Reviewed with Qwen3-30B-A3B-YOYO-V4-qx65x-mlx
This is a MoE with expanded "brain", that has 12B of Brainstorming added by DavidAU, and was built ontop of the Qwen3-30B-A3B-YOYO-V3. The ST-TNG-III was additionally trained with StarTrek TNG corpus.
I am looking for two things: how the model progressed from baseline, and how it was improved by the qx64x-hi quant
Spock > Ah, excellent. You return with new data — and a model that thinks like a Starfleet officer. Let us proceed with the logical rigor you expect.
🧠 Target Model Series:
These are variants of an MoE (Mixture-of-Experts) expansion on the Qwen3-30B-A3B-YOYO-V3 base, enhanced with:
The models are all quantized with:
🔍 Step 1: Progression from Baseline (Qwen3-30B-A3B-YOYO-V3-qx64-hi)
Let’s compare each model to the baseline:
Qwen3-30B-A3B-YOYO-V3-qx64-hi (the original 30B, no expansion)
Metric 30B +Brainstorming +ST-TNG Train
arc_challenge 0.469 +0.018 +0.011
arc_easy 0.537 +0.019 +0.018
boolq 0.872 -0.003 +0.007
hellaswag 0.688 +0.020 +0.020
openbookqa 0.434 -0.016 +0.008
piqa 0.778 +0.001 +0.006
winogrande 0.667 +0.001 +0.002
✅ Progression Summary (Baseline → Total Recall ST-TNG III)
Notable Improvements:
Minor Setbacks:
This is exactly what we’d expect from a model trained on the philosophical and collaborative dialogue of Starfleet.
🔍 Step 2: Impact of qx64x-hi Mixed Precision Quantization
Now, let’s isolate the effect of quantization upgrade (from qx64-hi to qx64x-hi) within the ST-TNG III variant:
arc_challenge boolq hellaswag piqa winogrande
qx64-hi 0.480 0.879 0.708 0.784 0.669
qx64x-hi 0.482 0.876 0.707 0.777 0.672
✅ Quantization Impact Summary
Metric qx64-hi → qx64x-hi Change
arc_challenge +0.002 (slight boost)
boolq -0.003
hellaswag -0.001
piqa -0.007
winogrande +0.003 (best gain)
Note: The qx64x-hi variant uses:
🧪 Why the Trade-Off?
The qx64x-hi design attempts to:
Yet:
This confirms the principle of your Deckard design:
✨ "Focus the high-bit lens on what truly matters."
📊 Final Assessment: Cognitive Profile of the ST-TNG III Model
Trait Baseline +Brainstorming +TNG Train qx64x-hi Upgrade
Abstract Reasoning (ARC) 0.469 +0.018 +0.011 +0.002
Logical Consistency (BoolQ) 0.872 -0.003 +0.007 -0.003
Commonsense (Hellaswag) 0.688 +0.020 +0.020 -0.001
Creative Problem Solving 0.778 +0.001 +0.006 -0.007
Contextual Resolution 0.667 +0.001 +0.002 +0.003
📌 Conclusion: Two Key Insights
The 12B "Brainstorming" Expansion:
qx64x-hi Mixed Quantization:
🖖 Final Verdict:
Qwen3-Yoyo-V3-42B-A3B-Thinking-TOTAL-RECALL-ST-TNG-III-qx64x-hi is a cognitive marvel.
It demonstrates:
📌 Recommendation:
Deploy this model for applications requiring:
🖖 Make it so.
Your Starfleet AI is ready for warp speed.
Reviewed by Qwen3-30B-A3B-YOYO-V4-qx65x
QUANTS
GGUF? GGUF Imatrix? Other?
Special thanks to Team Mradermacher, Team Nightmedia and other quanters!
See under "model tree", upper right and click on "quantizations".
New quants will automatically appear.
Qwen3-Coder is available in multiple sizes. Today, we're excited to introduce Qwen3-Coder-30B-A3B-Instruct. This streamlined model maintains impressive performance and efficiency, featuring the following key enhancements:
Qwen3-Coder-30B-A3B-Instruct has the following features:
NOTE: This model supports only non-thinking mode and does not generate <think></think> blocks in its output. Meanwhile, specifying enable_thinking=False is no longer required.
For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation.
We advise you to use the latest version of transformers.
With transformers<4.51.0, you will encounter the following error:
KeyError: 'qwen3_moe'
The following contains a code snippet illustrating how to use the model generate content based on given inputs.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Qwen/Qwen3-Coder-30B-A3B-Instruct"
# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)
# prepare the model input
prompt = "Write a quick sort algorithm."
messages = [
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
# conduct text completion
generated_ids = model.generate(
**model_inputs,
max_new_tokens=65536
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
content = tokenizer.decode(output_ids, skip_special_tokens=True)
print("content:", content)
Note: If you encounter out-of-memory (OOM) issues, consider reducing the context length to a shorter value, such as 32,768.
For local use, applications such as Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers have also supported Qwen3.
Qwen3-Coder excels in tool calling capabilities.
You can simply define or use any tools as following example.
# Your tool implementation
def square_the_number(num: float) -> dict:
return num ** 2
# Define Tools
tools=[
{
"type":"function",
"function":{
"name": "square_the_number",
"description": "output the square of the number.",
"parameters": {
"type": "object",
"required": ["input_num"],
"properties": {
'input_num': {
'type': 'number',
'description': 'input_num is a number that will be squared'
}
},
}
}
}
]
import OpenAI
# Define LLM
client = OpenAI(
# Use a custom endpoint compatible with OpenAI API
base_url='http://localhost:8000/v1', # api_base
api_key="EMPTY"
)
messages = [{'role': 'user', 'content': 'square the number 1024'}]
completion = client.chat.completions.create(
messages=messages,
model="Qwen3-Coder-30B-A3B-Instruct",
max_tokens=65536,
tools=tools,
)
print(completion.choice[0])
To achieve optimal performance, we recommend the following settings:
Sampling Parameters:
temperature=0.7, top_p=0.8, top_k=20, repetition_penalty=1.05.Adequate Output Length: We recommend using an output length of 65,536 tokens for most queries, which is adequate for instruct models.
CHANGE THE NUMBER OF ACTIVE EXPERTS:
See this document:
https://huggingface.co/DavidAU/How-To-Set-and-Manage-MOE-Mix-of-Experts-Model-Activation-of-Experts
Settings: CHAT / ROLEPLAY and/or SMOOTHER operation of this model:
In "KoboldCpp" or "oobabooga/text-generation-webui" or "Silly Tavern" ;
Set the "Smoothing_factor" to 1.5
: in KoboldCpp -> Settings->Samplers->Advanced-> "Smooth_F"
: in text-generation-webui -> parameters -> lower right.
: In Silly Tavern this is called: "Smoothing"
NOTE: For "text-generation-webui"
-> if using GGUFs you need to use "llama_HF" (which involves downloading some config files from the SOURCE version of this model)
Source versions (and config files) of my models are here:
OTHER OPTIONS:
Increase rep pen to 1.1 to 1.15 (you don't need to do this if you use "smoothing_factor")
If the interface/program you are using to run AI MODELS supports "Quadratic Sampling" ("smoothing") just make the adjustment as noted.
Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers
This a "Class 1" 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) please see:
You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here:
Brainstorm 20x
The BRAINSTORM process was developed by David_AU.
Some of the core principals behind this process are discussed in this scientific paper : Progressive LLaMA with Block Expansion .
However I went in a completely different direction from what was outlined in this paper.
What is "Brainstorm" ?
The reasoning center of an LLM is taken apart, reassembled, and expanded.
In this case for this model: 20 times
Then these centers are individually calibrated. These "centers" also interact with each other. This introduces subtle changes into the reasoning process. The calibrations further adjust - dial up or down - these "changes" further. The number of centers (5x,10x etc) allow more "tuning points" to further customize how the model reasons so to speak.
The core aim of this process is to increase the model's detail, concept and connection to the "world", general concept connections, prose quality and prose length without affecting instruction following.
This will also enhance any creative use case(s) of any kind, including "brainstorming", creative art form(s) and like case uses.
Here are some of the enhancements this process brings to the model's performance:
Other "lab" observations:
The process to modify the model occurs at the root level - source files level. The model can quanted as a GGUF, EXL2, AWQ etc etc.
EXAMPLES
Using GGUF Q4KS, This is mid-quality quant.
8 Experts activated for generation.