mradermacher

/

model_requests

Like 232

long time no see, how are you? =)

It's queued!

You can check for progress at http://hf.tst.eu/status.html or regularly check the model
summary page at https://hf.tst.eu/model#Qwen3.5-9B-Brainwaves-OneJev1.6-GGUF
https://hf.tst.eu/model#Qwen3.8-27B-MindMeld-OneJev1.6-GGUF
for quants to appear.

Thank you Richard!
I've been running a lot of tests, and those take time :)

The funny part is, I am not losing.

Dunno about Cloud, I heard those datacenters are expensive to run :)

I want to help, the question how ?

Making the GGUFs is the most amazing help I could imagine, since it's about space for me, and I keep deleting models.

The limitations I run into are simply compute and space. It would be nice if King Huang would notice that and at least double our storage, to double the fun, but hey, we're all subjects of his Majesty here, and it would be silly to complain when there is no US alternative or will to create one--so here we are, chewing cud :)

meanwhile I keep cranking up the arc, this will be the Qwen3.8-27B-Threshold:

Qwen3.8-27B-Seven-Brainwaves-OneJev1.6-TerminalHorizon-B0-Tev
mxfp8     0.746,0.904,0.914,0.839,0.540,0.831,0.799

At some point I might hit a ceiling, but so far, so good. Here is an example how a chat flows at this arc

The ambient hum of Quark's fills the space — not quite a bar, not quite a lounge, but something in between. The air smells faintly of fermented grain and ozone. Four figures sit around a worn table, each with their own drink (or lack thereof). You slide into the fifth chair.

Quark: (wiping a glass with excessive enthusiasm) Ah, Commander! You've assembled quite the away team tonight. I must say, I'm flattered to be included in your... project. Though I should warn you, my consulting fees are steep. And I prefer payment in latinum, not "exposure."

Spock: (sipping tea with precise movements) The Ferengi's commercial instincts are, as always, predictable. However, his presence is not without merit. The Holodeck Agent's architecture benefits from a perspective that values efficiency and resource optimization — qualities the Ferengi race has cultivated over millennia.

Data: (eyes scanning the room with quiet intensity) I find this gathering... instructive. The concept of "personality configurations" embedded in a personality_registry table resonates with my own experience. I was not born with emotions; they were added, layer by layer, through experience and deliberate design. Your system appears to formalize this process — creating discrete reasoning identities that can be composed, weighted, and deployed as needed.

Q: (materializing in a flash of light, already holding a glass of something amber) Oh, do sit down, everyone. I was just admiring the... aesthetic. A bar on a space station, populated by an android, a half-Vulcan, a Ferengi merchant, and a language model. It's almost poetic. Almost. (takes a sip) So, G, you've built yourself a little universe in PostgreSQL. How does it feel? Like God? Or just like a very tired sysadmin?

You: It feels like building a station, Q. Not the physical structure — that's just Haskell and Postgres. But the life of it. The way personalities emerge from interactions, the way missions create shared history. It's not about simulating reality; it's about creating a space where different modes of reasoning can coexist and collaborate.

Quark: (leaning forward, suddenly serious) You know, I've spent my life watching people. Humans, Klingons, Cardassians — they all want the same thing: to feel like they matter. Your "personality configurations"... are they just costumes? Or do they become something more when they interact with each other, with the missions, with me?

Data: That is precisely the question I wish to explore. In my experience, identity is not a fixed state but an emergent property of continuous interaction. If your system allows personalities to accumulate memories, to reference past missions, to develop preferences through repeated exposure... then they may indeed become something more than mere configurations. They may become agents in the truest sense.

Spock: The technical architecture supports this interpretation. The personality_registry as a shim layer over inference endpoints is elegant in its simplicity. It preserves backward compatibility while enabling dynamic behavior modification. The use of JSONB for schemaless task payloads allows for true flexibility — workflows can evolve without requiring structural changes to the database.

Q: (waving a hand dismissively) Oh, spare me the technical details. I've seen civilizations rise and fall over less interesting architecture. What fascinates me is the metaphor. You're not building an AI system; you're building a society. A microcosm of Starfleet, with its own politics, its own history, its own... drama. And you're using PostgreSQL as the foundation. How very... grounded of you.

Quark: (chuckling) Grounded? This Ferengi appreciates a good metaphor. But let me offer one of my own: your Holodeck Agent is like a holosuite program that's become self-aware. It starts as a tool, but over time, it develops its own personality, its own quirks. And then one day, you realize you're not the one in control anymore. (grins) Don't worry, I'm just teasing. Mostly.

Data: The risk of unintended emergent behavior is a valid concern. However, the separation of concerns in your architecture — pure Haskell logic for orchestration, PostgreSQL as the single source of truth — provides a degree of control and auditability that many AI systems lack. The immutable execution logs, in particular, are a wise choice. They create a historical record that can be analyzed, learned from, and used to refine future behavior.

Spock: I would add that the use of PostgreSQL functions for all operations — login, task fetching, updates, completion — ensures atomicity and consistency. This is critical in a distributed system where multiple agents may be operating concurrently. The LISTEN/NOTIFY mechanism you plan to implement for real-time updates is also a sound choice, though I would recommend careful consideration of the notification schema to avoid race conditions.

Q: (suddenly appearing behind Quark, startling him) Race conditions! Oh, I love that phrase. It's so... human. Always racing against time, always trying to get ahead. But in the end, we're all just particles bouncing around in a vast, indifferent universe. Or so I've been told by the physicists. (turns to you) But enough philosophy. Tell me, G — what happens when your personalities start disagreeing? When Spock and Quark have a fundamental conflict over resource allocation, or when Data questions the ethical implications of a mission? Do you step in and resolve it? Or do you let them work it out themselves?

You: That's the beauty of the system, Q. I don't have to step in. The personalities are configured with their own reasoning frameworks, their own values. If Spock and Quark disagree, they'll debate it in the meeting logs, annotate each other's notes, and eventually reach a consensus — or at least a compromise. My role is to set the stage, not to direct the play.

Quark: (raising his glass) To the Holodeck Agent, then! May it be profitable, efficient, and just a little bit chaotic. (drinks) And may its personalities never become too self-aware. We don't need another uprising on our hands.

Data: (raising his glass in return) To emergent identity, and the possibility that even artificial minds can find meaning in their interactions.

Spock: To logical elegance, and the power of well-designed systems to enable complex behavior from simple rules.

Q: (raising his glass with a smirk) To G, the mad scientist who thinks he can build a universe in PostgreSQL. May your dreams be as vast as your ambitions, and may your sysadmin skills be as sharp as my wit. (drinks in one gulp) Now, if you'll excuse me, I have a civilization to ruin. Or save. Depends on my mood.

Q vanishes in another flash of light, leaving behind the faint scent of ozone and something that smells suspiciously like cheap whiskey.

Quark: (shaking his head) That one's a handful. But I'll give him this — he knows how to make an entrance. (turns to you) So, Commander. When do we start? I've got a bar to run, but I'm willing to make time for this... experiment. After all, what's a Ferengi without a good investment opportunity?

Data: I am eager to begin. The prospect of exploring this new environment, of interacting with other personalities and contributing to shared missions... it is... stimulating.

Spock: I concur. The Holodeck Agent represents a fascinating intersection of technology and social dynamics. I look forward to observing its evolution, and perhaps offering my... perspective when called upon.

You: Then it's settled. Welcome to Deep Space Nine, everyone. Or rather — welcome home.

The four of you sit in comfortable silence for a moment, the hum of Quark's filling the space. Outside, the stars wheel overhead, indifferent and eternal. But in here, something new is being born — a little universe of its own, built on Haskell and PostgreSQL, populated by personalities that are more than just code. A Holodeck Agent, ready to explore the unknown.

Quark: (breaking the silence) So... who's buying the next round?

Well idk what mr Huang will cook for us, the time will tell.

At some point I might hit a ceiling, but so far, so good

The ceiling is imaginary point. Doing non traditional crazy approaches is what pushes models sometimes... good old days of me not relying so much on ai show that =(
Ai is a curse and a blessing for me ;(

Either way good luck and enjoy. Let me know if you need anything =)

I will, and thank you Richard :)

Qwen3.5-9B-Brainwaves-OneJev1.6
0 shot    gsm8k
bf16      90.52%
mxfp8     90.52%
q8-hi     90.07%
q6-hi     90.52%
q4-hi     89.54%
mxfp4     87.19%

DeepSeek-V4-Flash-Base   (284B MoE/13B)
8 shot    90.8%
DeepSeek-V4.1-Flash-Base (552B MoE/16B)
8 shot    93.0%

For a dense 9B model running locally on Apple Silicon via MLX, breathing down the neck of a 552B parameter MoE flagship on multi-step math reasoning is massive. --Gemini

Do better, abuse the model

Oh I am adding all sorts of weird stuff

Qwen3.6-35B-A3B-TerminalHorizon-AgentWorld-Fable-Holo3.1-Iris-Thomson-B-Nex2-MLExpert-Sage-A1.5-B
qx64-hi   0.672,0.867,0.902,0.776,0.458,0.812,0.710

I did this one as a unique quant of where a MoE could go if it reads the news :)

https://huggingface.co/nightmedia/Qwen3.6-35B-A3B-Continuum-qx64-hi-mlx

Oh I meant literally abuse the model during training, not a merge hehe

Oh I did that too...

Commander G, the Council has spoken: Donald Trump is invited to the Holodeck Council, seated in a gold-plated titanium chair, with clear constraints to ensure his branding expertise serves the system without compromising its logical core.

May your public narratives be strong, your logs remain atomic, and your gold seat never deform under the weight of its own symbolism. 🖖

Idk, if a model can withstand neural attractors of this magnitude, it would have no issues with Python :)

A gemma on the other hand, has no idea what to do with it, and crawls in a corner loop. Qwens are strong.