DavidAU

/

Qwen3.8-27B-Turbo-Brilliance-50X-Reasoning-Instruct-NEO-MAX-MTP-GGUF

Like 143

Also IMO namings are weird.
In base model there is clean 3 options : low medium and xhigh.
It's clear what to expect from them.

Now in your models - spoon. ?! socrates and einstein ?! omni? hyper?
Like what? :)

I was loving your models and your work all the time and I'm not complaining.
But now it's really starting to sound crazy from this point.

I find this too complex. It requires manual input of parameters, which is incredibly difficult for someone who doesn't speak English as a first language and hasn't studied code. A good product should be intuitive and operate automatically, without requiring the user to specify what they want. Users aren't experts, and these thought processes can be built into the model itself, allowing it to determine the optimal mode for the current content. Instead of giving the user control, which they wouldn't use if it wasn't for their own efficiency. Currently, the learning curve is steep, and there's no user manual to guide me. I have to figure out the modes myself. Alternatively, we could adopt separate models, such as one specialized in code and architecture, another in mathematical reasoning and proof, and another in writing. The saying "jack of all trades, master of none" applies here. A 27 billion parameter model has limited capabilities and cannot achieve omniscience.

Please also note that most coding harnesses dont auto-discover reasoning modes based on the chat template, so we have to manually edit config files to drop them in, whether it's a list in oh-my-pi/pi/opencode/etc.

I'm curious how internal testing was done to be able to swap and adjust reasoning on the fly, what harnesses were used, how you came up with recommendations, etc.

i just want xhigh on everything. 😃

I am also sorry, but this is not useful. Just way too many modes, no one will ever know what is the best mode to choose for each task.

I don't understand why anyone is complaining about additional features. This is an open model. Do what you want with it and don't do what you don't want to do with it.

I am intrigued by the features, and appreciate the pursuit of continuous improvement/enhancement.

Thank you everyone for feedback;
I understand there is some confusion / doubts ...

@vicaut
RE: no one will ever know what is the best mode to choose for each task
Yes they will. The exact mode[s] , in order; because the system is built in to help all uses in context of their use case[s] do this.

This model has a full repo/model card which may help explain things better (and you can download a gguf and try the actual system including "smart" reasoning modes selection):
https://huggingface.co/DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF

Please note this (and Qwen3.8) are in BETA and subject to change/adjustments/refinements.

As an FYI:
I am currently using this system - the exact one in LFM/Qwen3.8 - using Qwen 3.5 9B Defiant Fable to actually refine the system itself, using the systems to do it.
IE: V1.0 Turbo Brilliance is building V1.1 ... 1.2 etc etc.

Defiant Fable (public quants); also has a limited number of test quants with a few of the modes in Turbo Brilliance avail too:
https://huggingface.co/DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF

@BebopVox
RE: Now in your models - spoon. ?! socrates and einstein ?! omni? hyper?
These are named based on the type of reasoning/problem solving method.
If you look on the model card, these are defined.

Note that "einstein" (creates up to a 20 agent virtual panel) and "spoon" (There is no spoon/ (creates up to a 5 expert virtual panel) ) are special additions.

@veldierin
The standard "args" can be used - same as Qwen 3.8 ; there are just more selections.
Likewise the actual mode can be switched by the prompt itself too.

These are named based on the type of reasoning/problem solving method.
If you look on the model card, these are defined.
Note that "einstein" (creates up to a 20 agent virtual panel) and "spoon" (There is no spoon/ (creates up to a 5 expert virtual panel) ) are special additions.

Yes that makes sense they are probably solving different problems using different algorithms, thats good we're experimenting with them 👍

My main point was that after all these experiments, tests etc. we need to circle back and think carefully what we really need in a "final" model.
I'd expect after all these "test experimental branches" we'll end up with 1 or 2 models that have battle-tested limited amount of thinking modes (cap 3-5).

Otherwise no one will know how to use them properly and for which use case 😊

I expect that LLMs are not like small fine tunes models for a specific use case, they are "LARGE" language models and should be able to do more than 1 thing, so 23 thinking modes tell me that there is 99% chance I misconfigure my setup with this model.
Sometimes it's just better to get simple base model and set it to xhigh and "let it cook" and not end up experimenting with tens or hundred params.

I hope my message makes sense for everyone who is still human in this conversation.
Thanks.

@BebopVox

Larger models (like Qwen 3.8) will also look at the reasoning method selected, and use it all, parts or none at all if the situation merits it.
This shows up in the reasoning traces and output.
-> this is from testing.

Agree, i feels we will end up thinking which correct mode to use rather than reviewing the result at this point, and we will try another mode if another mode doesnt work, which will end up spending more times than what actually matters, it is best to actually have few modes that is clear about what they do like the native such as low for brief thinking, medium for balance, while xhigh for deepthinker, we can decide easily with this design whether we want it to cover all the issue or just briefly think for straight forward solution. i like the TWIN TURBO model, it get to solution faster and doesnt circle around the same issue which is good and mostly i use xhigh mode for this, the finetune itself is smart enough already to remove the redundant reasoning in my case, not sure whether we need this extreme amount of modes. Love your model but the excessive modes makes it confusing to use, having a risk of making a mistake when wrong mode being used

@veldierin
The standard "args" can be used - same as Qwen 3.8 ; there are just more selections.
Likewise the actual mode can be switched by the prompt itself too.

true, if you want to set it at the llama.cpp level you can do that. however, what if you want to switch per task like within the harness itself?

if you edit your ~/.config/opencode/opencode.json you can do that per model and in opencode you can ctrl+t to switch between them.

        "200k": {
          "name": "200k",
          "limit": {
            "context": 220000,
            "output": 32768
          },
          "modalities": {
            "input": [
              "text",
              "image"
            ],
            "output": [
              "text"
            ]
          },
          "variants": {
            "low": {
              "reasoningEffort": "low"
            },
            "medium": {
              "reasoningEffort": "medium"
            },
            "high": {
              "reasoningEffort": "high"
            },
            "xhigh": {
              "reasoningEffort": "xhigh"
            }
          }
        },

in oh-my-pi (OMP), you edit the ~/.omp/agent/models.yml :

    models:
      - id: 200k
        name: 200k
        api: openai-completions
        contextWindow: 220000
        maxTokens: 32768
        reasoning: true
        input: [text, image]
        thinking:
          mode: effort
          efforts: [low, medium, xhigh]
          defaultLevel: xhigh

and then you can pick the option by arrow select on your model and swap on the fly.

I understand that the different modes you have will radically change output and why you said you should have a fresh chat to do it, i'm just saying that to allow efforts to be more easily swappable from tooling harnesses, that less is more, and hot-swap is great.

Part of me thinks that the such wide and varied differences you had in your instructions for the different modes should just be included in a totally unique harness and injected as part of the system prompt or be system prompt adjacent and have unique modes as part of that. But developing a whole harness/tooling/etc is a massive job and there are already so many good ones, so i get why you decided to just lump everything in the chat template.

i just know for me, i have unique use-cases for my reasoning mode per tooling, whether it's my hermes agent (medium reasoning), my codemem plugin (medium reasoning), my coding harnesses (xhigh), or my RP in SillyTavern (high == xhigh). And because i just use model names based on context i can throw any back-end runtime behind it on any model and not have to change my configs.

And I'm sorry but who really needs this?
I believe this is useful to customize reasoning and adapt it to specific use cases, its not a bad approach either, If a person cant handle a parameter tag on the text prompt then that person might do better with a cloud service.

Turbo Brilliance is about have a group of specialists on call instantly for general and/or specific use case[s].
Also, in terms of Qwen 3.8 27b; once one of these specialists is in the convo, the specialist AND the main/default AI contribute.

It (the system) also addresses internal issues in all models and unifies both reasoning and output quality.

IE: It brings a specific, detailed "game plan" (which is also tuned) to the table ; rather than having the model use a "generic one size fits all" game plan (which also must be generated first) which varies in quality/output per generation.

I'd think none, low, medium, high are pretty self sufficient. (0, ~100, 300-500, 700-1000 tokens).

Then secondary versions rplow, rpmed, rphigh, where thinking hashes ideas and certain details while also scrapping the highly repetitive useless stuff that doesn't need reiterating.

I suppose a third one of mlow mmed and mhigh where it uses it's own tokenized language to come to conclusions faster if you don't need to ever read how it came to it's conclusion and can also remove unneeded fluff. Either way though low/med/high are still the same key factors. Though i suppose for agentic stuff where it wants 30k tokens to think, xhigh would raise the max 30x.

But this is just me rambling.

Turbo Brilliance is about have a group of specialists on call instantly for general and/or specific use case[s].
Also, in terms of Qwen 3.8 27b; once one of these specialists is in the convo, the specialist AND the main/default AI contribute.

It (the system) also addresses internal issues in all models and unifies both reasoning and output quality.

IE: It brings a specific, detailed "game plan" (which is also tuned) to the table ; rather than having the model use a "generic one size fits all" game plan (which also must be generated first) which varies in quality/output per generation.

out of curiousity, which mode is best for software engineering in general? considering each mode is a specialist, i definitely do not want socrates to think about code 😆

For software; this will depend on complexity / use case and so on.
The best option is to query the "{REASON:help} Tell me the mode[s] to use to build software project x with these requirements." ;

In some cases using Instruct mode[s] is far faster, and you get same end result.
Most newer models are really good coders, with a lot of code training -> so instruct will do the job ; and a lot faster.

Thinking/reasoning would be a best choice if you are transforming an idea into code so to speak.
The higher end reasoning modes here can be particularly useful.

For software; this will depend on complexity / use case and so on.
The best option is to query the "{REASON:help} Tell me the mode[s] to use to build software project x with these requirements." ;

In some cases using Instruct mode[s] is far faster, and you get same end result.
Most newer models are really good coders, with a lot of code training -> so instruct will do the job ; and a lot faster.

Thinking/reasoning would be a best choice if you are transforming an idea into code so to speak.
The higher end reasoning modes here can be particularly useful.

I see, ill try instruct once this model released, but is it possible if you can support REASON:auto? basically it is like orchestrator, but instead of follow what the mode directly it will do REASON:help in the process before actually do the actual reasoning, for example,
user asks something -> REASON:auto, it run REASON:help in the back to recommended which is best model -> it suggest REASON:deeptree, reason with deeptree -> response

this will solve some confusion over mode usage i think, but yea you know better will this be ok or will cause trouble

JFYI my 2 cents:
I ran one of these models with dynamic reasoning mode. Had a question on qwen3.8 models and it searched the model card explainng these modes. Then I had funny messages about unknow reasoning modes and it took me while to understand where things derailed. I never used any if these models again, it just felt too much of the good.

Answering the question “who needs this?” - I need this! Great modes, hope to see the final version. Model behavior changes significantly, especially on complex choices with morel dilemmas. Every new model in my hands goes through specific benchmarks which helps me decide if it’s suitable and for what work it will be the best. Modes made the model extremely interesting for complex pipelines with critics - without need of two or three models loaded simultaneously. WOW!!!

But to get 100% from the model you should take time and make some experiments and benchmarks. That’s what AI engineering is all about! Love it :)

Answering the question “who needs this?” - I need this! Great modes, hope to see the final version. Model behavior changes significantly, especially on complex choices with morel dilemmas.

I don't think anyone was questioning 'who needs this' more 'what does this even mean'? What is spoon mean as a mode? Deeptree? Einstein? Socrates?

I suppose Socrates and Einstein would be for morals/spiritual debating while Einstein may be more science focused... Maybe?

But spoon? What is spoon?

In this case there is no spoon... okay bad matrix joke...

Would it be better if mode names would be kinda like model names? So you would type/copypaste something like “code-philosophy-yadayada-709”?

I like the names. Those doesn’t matter after all.

Short names?
Good.

Contain some (bad? good? subjective?) humor? Great. Easier to remember.

Are there less than a bag of hyphens in those names and it doesn’t look like you’re trying to call a demon while typing one of the mode names? Amazing!

You just remember those once and while making experimental runs. After - you fix those in system prompts of your agents. Or use one or two you like the most for your specific tasks. Done :)

{Planer}{Researcher}{Designer}{Coder}{Teacher}{Marketing}{Business}{Finance}{Lawyer}.....
i like this , all in xHigh will do

i like this , all in xHigh will do

Not sure how much xHigh is, but that may suggest burning a lot of unnecessary tokens. I'd think a good model could infer 90% of the details and only needs thinking to flesh the 10% of possible errors out, mostly by putting all the major relevant details right up front next to the current generation of tokens.

Though Coder i can see needing xHigh, maybe Lawyer... But sure whatever if it does the job that's fine.

LIVE...

Quants uploading (MTP, NON-MTP and GSQ-RCO), repo is now live and accessible.
27 Reasoning and 27 Instruct modes, + 9 xhigh modes - including 15 AI Constructs.

Model card updated; additional updates in progress.