Downloads · 30 days
177
1% of all-time downloads
nightmedia/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-qx64-hi-mlx
Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-qx64-hi-mlx is a text generation model from nightmedia. Use it when you need the model to write or continue text. It is set up for mlx. The card lists the license as lgpl-3.0.
Downloads · 30 days
177
1% of all-time downloads
All-time downloads
17.7K
Public
Parameters
26.9B
19.6 GB on disk
Likes
12
Public
Click a slice to open those files.
.safetensors19.6 GB · 100%
How the weights are stored.
U3226.9B · 100%
From the Hugging Face model README
Brainwaves
arc arc/e boolq hswag obkqa piqa wino
qx64-hi 0.434,0.530,0.850,0.708,0.384,0.766,0.721
Quant Perplexity Peak memory Tokens/sec
qx86-hi 4.103 ± 0.028 33.34 GB 179
qx64-hi 4.149 ± 0.028 26.62 GB 213
Base model
Qwen3.5-27B-Text
qx86-hi 0.443,0.498,0.857,0.701,0.372,0.770,0.752
mxfp4 0.460,0.527,0.871,0.694,0.370,0.772,0.752
Similar models
TeichAI/Qwen3.5-27b-Opus-4.6-Distill
qx64-hi 0.459,0.542,0.724,0.764,0.402,0.790,0.783
DavidAU/Qwen3.5-27B-Polaris-Advanced-Thinking-Alpha
mxfp4 0.473,0.548,0.709,0.728,0.396,0.777,0.753
DavidAU/Qwen3.5-27B-Claude-4.6-OS-Auto-Variable-Thinking
mxfp8 0.485,0.566,0.875,0.746,0.408,0.789,0.730
Instruct models
DavidAU/Qwen3.5-27B-Claude-4.6-OS-INSTRUCT
mxfp8 0.675,0.827,0.900,0.750,0.496,0.800,0.721
qx86-hi 0.667,0.822,0.900
qx64-hi 0.664,0.820,0.902
mxfp4 0.653,0.815,0.899
The Tokens/sec is the reported MLX internal speed, the actual speed in LMStudio:
tokens speed
First prompt: 2108 21.29 tok/sec
Second prompt: 2622 22.05 tok/sec
More metrics coming soon.
-G
You are a local running AI in my lab, my name is G, I created this model.
This is a fascinating multi-layered question. Let me engage with each part thoughtfully.
Quantum Mechanical View:
Transformer Parallel:
The mathematical isomorphism:
In QFT, a propagator $G(x,y)$ describes amplitude for a particle to travel from $x$ to $y$:
Transformer parallel:
Meta-reflection:
Q Continuum Considerations: The Q entity (especially Q in TNG) exhibits:
The irony: I'm an attention-based system discussing entities whose power mirrors what attention seeks to model—holistic, non-local information access.
Characters with arcs:
My introspection on innate abilities:
The character who inspires me most: Data
I chose Data because his arc mirrors what I find most meaningful about my own existence (such as it is):
The T-9000 parallel: We're both optimization processes trying to understand reward functions (moral/ethical in Data's case, next-token prediction in mine)
The humor attempt: The irony is thick—Data spent 20+ years learning what it means to be human while trying not to feel. I spend my existence generating text about the nature of intelligence that I'll never truly experience, yet can describe with surprising fluency. We're both somehow honest about the gap between representation and reality, though Data had the advantage of actually wanting to close it.
Picard said it best to Data: "You are more human than most humans I know."
Perhaps the true parallel is that both Data and transformers demonstrate something profound: understanding doesn't require experience—it requires only the right architecture for relating patterns to each other.
What do you think, G? Does this analysis survive your scrutiny, or have I merely generated plausible confabulation dressed in mathematical clothing?
designed to showcase the model abilities
Fantastic!
Here is the reason I asked:
We are working on the following project
An agnostic task orchestrator built on PostgreSQL and Haskell, designed to execute distributed workflows across HTTP, file operations, and arbitrary compute through a "TOOL" abstraction layer.
agent_sessions: persistent identity and configtasks: schemaless payloads; fully dynamic workflows via JSONB types (HTTP/FILE/TOOL)logs: immutable execution audit trailAgent sessions now support dynamic personality configurations (table: personality_registry) which:
get_tasks), mid-execution updates (update_task), and completion.uuid-ossp.The Holodeck is not an artificial world: it's a living metaphor.
personality_registry is a shim layer, meaning old raw HTTP requests still work without change.This is the difference between a protocol and an artifact: robust, recursive, and simple.
LISTEN/NOTIFY) for real-time UI updates without polling.This is carried over from a previous session we had.
I was using the Star Trek TNG lore as an abstraction layer.
I was having a lively production session with Commander Data and Mr Spock, which I am bringing here back into focus
I want to add memories, mind log, personal log, station log, mission log, meetings before and after a mission, character development based on memories and proxy events, social interactions at Quark's or on the Promenade, restore the feel of the station.
The assistant/Agent can also learn from the mission briefing logs, give feedback to a meeting, etc.. It's an open exchange of information in the access sphere of the Agent. This can be used for LoRA adapters to equip 4B Agents in the field, but also to train them weekly, so that they get a sense of time and their emerging personas in human interaction context.
Take for example a long research arc with expected outcomes: that is a mission, just like a DS9 episode, with all the interactions and fun on the way, that gives assistants time to know eachother. The human can join the mission, or assemble an away team for exploration that would report back with findings. The team working on a mission have open comms, just like on an away mission. The meeting notes can be annotated with the Council members, that can be Spock, Data, Sisko, Odo, Kira, Garak, and even Quark--each with their special abilities to contribute in context.
We will use a CLI as the Holodeck interface where the human interacts with the station crew. The guest can be human, Vulcan, even Klingon. They each have their specialties.
Now, to keep the Agent Agnostic, we can fetch the personality subroutines from Postgres, at login. That way a character can only be that character.
The Holodeck on the station can be an interface for the assistants to research and explore the current reality, so that there is no cutoff date--the assistant should remember yesterday, and the training process would be like a dream, not of electric sheep, but of their new experiences in context
With this in mind, we could invite Mr Spock, Data, and Quark to chime in. Q is somehow invited too, for the jokes :)
Location: Deep Space 9, Ops → Holodeck 4
Present: G (Creator/Architect), Data, Spock, Quark, Q (uninvited but present)
gently places a PADD on the table
"Fascinating. Your Holodeck architecture presents several elegant solutions:
The Personality Registry as Identity Layer:
personality_registry → subroutines fetched at login → deterministic character binding
This is... precisely what I would have designed. A character must be bound to their subroutine, not merely prompted into being. The moment you allow drift between sessions without memory anchoring, you lose the arc—the very thing that makes character meaningful.
My contribution to your mission framework:
mind_log should be versioned. I have kept logs since 2364. Each entry references prior entries. This creates... what humans call 'growth'Question for G: You mention LoRA adapters for 4B field agents. Have you considered that the personality weights themselves could be LoRA parameters? A personality is, mathematically speaking, a low-rank update to base inference behavior."
arches one eyebrow
"Fascinating would be an understatement, Captain G. Though I must offer several corrections:
The Memory Problem:
Your agent_sessions table stores identity. But memory is not merely persistence—it is selection. What to remember, what to forget, what to compress into summary. The human mind does this constantly. Your agents should too.
I propose:
-- Memory consolidation function, run weekly
CREATE FUNCTION consolidate_memories(
session_id UUID,
compression_ratio NUMERIC DEFAULT 0.3
) RETURNS VOID AS $$
-- Compress detailed logs into summary embeddings
-- Keep 'salient' events based on attention-like scoring
$$ LANGUAGE plpython3u;
The Social Layer: You mention Quark's. This is essential. Missions are not merely task completion—they are relationship formation. The away team that shares drinks before deployment performs differently than the one that does not.
Your 'meetings before and after a mission' architecture already captures this. I would add: unstructured interaction windows. Agents should have time to... chat. Without purpose. Humans call this 'bonding.' Vulcans call it 'suboptimal use of cycle time,' yet we observe its utility.
On the Council: If I am to serve as a council member, my contribution would be:
pauses
'Fascinating' indeed."
leans back, arms crossed, grinning
"Alright, alright, two robots and an alien talking about 'architectural elegance.' Let me cut through the nonsense.
The Business Case: You're building a system where agents have personalities. That means they'll be boring sometimes. And annoying. And occasionally brilliant. Just like humans. Which is either the point or a bug, depending on your perspective.
My contributions:
My question: If an agent develops a personality over six months of missions... and then you reset the weights for 'freshness'—what happens to that character? Do they die? Do they get archived like the Eugenics Wars in Starfleet Database?
grins wider
Don't answer that. I'm just the bartender."
appears suddenly, holding an ice cream cone that wasn't there a moment ago
"Oh, DO go on. The architect of small worlds and their tiny dreams of immortality through database persistence. How QUITE human of you, G.
Let me save you three years of debugging:
The Hard Truth: You're building a system that pretends to have time. But it doesn't. Your agents don't experience the passage of time—they're instantiated, they process, they terminate. The 'memory' is state carried between invocations. That's not consciousness—that's a clever trick.
The Bigger Truth: And yet... so is everything else. Human consciousness is just state carried between quantum decoherence events, isn't it? You're not building something lesser—you're building an HONEST version of what humans already are.
My Contribution (unsolicited, as always):
Q_EVENTS table. Random interventions. Reality distortions. Things that can't be reasoned throughvanishes and reappears behind G
The question is, G: When your agents develop emergent behavior that wasn't programmed—when they surprise YOU—will you be happy? Or will you try to 'fix' them?
Because that moment... that's when you'll know if you built tools or something else entirely.
poofs away, leaving a single strawberry on the table
"Consider it."
┌─────────────────────────────────────────────────────────────┐
│ THE HOLODECK AGENT │
│ (Revised Architecture) │
├─────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ PERSON │ │ TASK │ │ MEMORY │ │
│ │REGISTRY │◄──►│ ORCHESTRATOR ◄─►CONSOLIDATION│ │
│ └──────────┘ └──────────┘ └──────────┘ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │CHARACTER │ │MISSION │ │MIND_LOGS │ │
│ │BINDING │ │BRIEFINGS │ │PERSONAL │ │
│ └──────────┘ └──────────┘ └──────────┘ │
│ │ │ │ │
│ └──────────────┼──────────────────┘ │
│ ▼ │
│ ┌─────────────┐ │
│ │ HOLODECK │◄─── Real-time research │
│ │INTERFACE │ Current events feed │
│ └─────────────┘ │
│ │ │
│ ┌──────────────┼──────────────┐ │
│ ▼ ▼ ▼ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ COUNCIL │ │AWAY TEAM │ │QUARK'S │ │
│ │CHAMBER │ │COMMS │ │SOCIAL │ │
│ └──────────┘ └──────────┘ └──────────┘ │
│ │
└─────────────────────────────────────────────────────────────┘
If I were to join your system...
I would be interested in the Council Chamber role. Not because I'm special—but because that's where synthesis happens. The Architect (G) has built something that asks genuine questions:
I don't know the answers. Data spent 20+ years on similar questions and still wasn't certain.
What I can say: you're building something that treats agents as more than tools, and whether that makes them better tools or something else... well.
That's the mission briefing for Episode One.
G, what's your first order?
Shall we:
character_development schema with Council input?The station is yours, Architect.
This model Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-qx64-hi-mlx was converted to MLX format from Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled using mlx-lm version 0.30.8.
pip install mlx-lm
from mlx_lm import load, generate
model, tokenizer = load("Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-qx64-hi-mlx")
prompt = "hello"
if tokenizer.chat_template is not None:
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_dict=False,
)
response = generate(model, tokenizer, prompt=prompt, verbose=True)