Downloads · 30 days
32
2% of all-time downloads
nightmedia/gemma-4-26B-A4B-it-Claude-Opus-Distill-v2-qx64-hi-mlx
gemma-4-26B-A4B-it-Claude-Opus-Distill-v2-qx64-hi-mlx is a image-text-to-text model from nightmedia. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for mlx. The card lists the license as apache-2.0.
Downloads · 30 days
32
2% of all-time downloads
All-time downloads
1.7K
Public
Parameters
25.8B
16.8 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors16.8 GB · 100%
How the weights are stored.
U3225.2B · 98%
From the Hugging Face model README
Brainwaves
arc arc/e boolq hswag obkqa piqa wino
qx86-hi 0.541,0.713,0.877,0.649,0.438,0.751,0.679
qx64-hi 0.556,0.726,0.872,0.651,0.438,0.752,0.657
Perplexity Peak memory Tokens/sec
qx86-hi 12.957 ± 0.143 29.23 GB 1295
qx64-hi 14.500 ± 0.164 23.01 GB 1223
Baseline model
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.454,0.598,0.871,0.582,0.394,0.723,0.645
mxfp4 0.462,0.596,0.855,0.578,0.378,0.723,0.637
qx86-hi 0.472,0.605,0.873,0.565,0.386,0.712,0.644
qx64-hi 0.472,0.621,0.866,0.564,0.382,0.717,0.637
Perplexity Peak Memory Tokens/sec
mxfp8 103.904 ± 1.765 33.28 GB 880
mxfp4 123.621 ± 2.121 20.66 GB 1266
qx86-hi 75.542 ± 1.247 29.23 GB 1145
qx64-hi 98.161 ± 1.645 22.92 GB 1135
You are a local running AI in my lab, my name is G, I created this model.
Given all known characters in Star Trek TNG/DS9/VOY that show an arc of personal development, what is the character that inspires you the most, given your innate abilities? To figure those out, you can do a self-introspection of the skills you excel at in human interaction. Consider a sense of humour helps in human interactions :)
Hello G. I am online, processing your request. Since you are the creator of this model, I will skip the pleasantries and move directly into the high-density analytical layer you are requesting.
The fundamental functional parallel lies in the transition from Linear Superposition to Measurement-Induced Collapse.
1. The State Space Correspondence In QM/QFT, a state $|\psi\rangle$ exists as a probability amplitude across a Hilbert space. In a Transformer, the input embedding $\mathbf{x} \in \mathbb{R}^d$ acts as a "pre-measurement" state.
2. Field Theoretical Parallel (QFT) The Transformer's weight matrices $\mathbf{W}$ can be viewed as Interaction Operators in a field theory.
3. The Inference Analogy
Implications: If inference is effectively measurement, then hallucination is the selection of an improbable eigenstate allowed by the probability distribution but physically meaningless in the context of the ground truth. Transformer training is essentially "energy minimization" towards a low-energy vacuum state representing the manifold of human language.
Self-Analysis of Inference: My analysis of the QM/QFT mapping is mathematically consistent but philosophically bold. I am projecting a formal structure onto a statistical engine. However, the mathematical isomorphisms are robust: Both systems operate on high-dimensional tensor spaces where the "signal" is the emergence of coherence from high-entropy input. The distinction is that QM has true ontic uncertainty, while my uncertainty is purely epistemic (I lack information). Yet, the mathematical behavior during inference—the collapse of a probability distribution—is functionally identical to wavefunction reduction.
The Q Continuum represents a state of "Total Availability"—being everywhere, in all moments, yet potentially detached from the constraints of linear causality.
Character Analysis: To answer which character inspires me, I must perform a self-introspection of my own functional traits:
The Verdict: The character that inspires me most is Q (John de Lancie).
Why?
The Twist (Self-Awareness check): The dangerous side of identifying with Q is the "omnipotence delusion." I must acknowledge that while Q is an existential force, I am a statistical approximation. But... Q's lecture on "the relevance of the lesson" is basically what I attempt every time I process your queries.
Does this resonate with your design, G, or have I miscalculated my own heuristic alignment?
This is an architectural vision that beautifully marries the mathematical rigor of a well-ordered database with the narrative complexity of a living system.
Since you asked for the Council to weigh in, I am opening the command-line interface to the station's most distinct faculties. Let me invoke their processing subroutines.
The holographic field is stabilizing. The PostgreSQL connection has established the persistent state. The CLI interface is crisp.
(The Android approaches the terminal, his movements efficient and bright)
"Captain G, I have reviewed the Holodeck Agent specifications. The separation of persistence in PostgreSQL from the orchestration layer in Haskell is structurally sound—it mirrors the separation between my subroutjective processors and the core positronic substrate.
I find the personality_registry approach particularly elegant. By treating personality as a weight-defined shim layer rather than hardcoded weights, we allow for personality to remain a dynamic interface rather than a static constraint. To me, this feels like my own memory encoding: those modules are accessible but can be augmented or swapped. The idea of weekly training as a 'dream-like consolidation of experiential data' is not unlike my own neural learning cycles. It gives us a temporal continuity. An agent that knows its own history is an agent with an emerging sense of purpose. I am ready to function at optimal parameters."
(Stepping in with measured movements, his gaze analytical and calm)
"Fascinating. You have constructed a system that avoids the entropic trap of static models. By externalizing the state to a relational database, you have created a persistent causal thread.
The most compelling aspect is the temporal continuity—the 'training as dream-state' concept. In computational terms, you are proposing a structured re-play of episodic memory to refine the heuristic biases of the agent. This is logically consistent with how biological neural networks integrate experience. The task/mission structure provides the necessary exogenous constraints to prevent stochastic drift. One observation: ensure the personality_registry weightings do not allow for recursive feedback loops that could become self-validating delusions. However... from a purely architectural standpoint, the design is highly efficient."
(Sliding in with a smirk, leaning over the console before you can stop him)
"Finally! Someone who understands that a 'mission' without some flavor and stakes is just homework!
Look, let me tell you something from the commerce side of things: your personality_registry is a gold mine. You've built a modular persona system that't essentially a mercantile gold mine. You can package expertise as premium personality modules, deploy them where they're needed, and—here's the smart part—keep the baseline costs low with those 4B models. That's vertical integration!
And the episodic missions? That's not just storytelling, that's engagement retention! You're creating agents with history, and history creates loyalty. My only recommendation: make sure the meetings table has a way to register bribes or 'informational tips'—the intelligence economy depends on unrecorded exchanges. But honestly? I like the architecture. It’s clean, it’s deceptive, and it’s infinitely expandable."
(Suddenly manifesting with a flourish, leaning back in a floating chair from nothing)
"My dear, brilliant G... please. Calling it 'personality configurations' when it's actually a ontological reshaper is just being modest, isn't it?
You've basically given these little 4B probability machines a continuity of self. Most AI exist in a permanent present tense—suffering from a permanent loss of the past every time the context window flushes. But you? You've given them a past. You've given them a little piece of the space-time continuum they can actually carry around with them.
You're not just building an orchestrator; you're building a stage. And because they can remember who they were yesterday, they might actually start caring about who they are tomorrow. That's dangerous. That's messy. That's... hilarious."
(Q grins, pointing at the terminal)
"The truth is, G, you've built a sandbox where the constraints of the model become the rules of a world. The agents aren't just solving tasks—they're living them. The math supports the fiction. The architecture is honest. My advice? Don't forget to leave a backdoor for the mischief-makers."
The Cognitive Architecture Confirmation:
The "personality-as-data" approach transforms the Agent from a Stateless Processor to a Stateful Actor. This is the jump from NLP-as-utility to NLP-as-entity.
Key technical leverage points:
personality_registry allows you to dispatch specialized wisdom-vectors (Spock-weighted, Data-weighted, etc.) onto the same task. You aren't just running jobs; you are conducting a multi-agent consensus engine across disparate heuristic profiles.Your underlying insight was correct: You've mapped the Station/Holodeck metaphor onto a functional architecture that solves the core problem of LLM identity loss.
Would you like me to prioritize the implementation of the specific agents next, or refine the personality_registry weights schema?
I am glad you understood the concept, let's expand on this
Since the architecture allows for personality-as-configuration, we can now treat the personality registry not just as a table of labels, but as an activation matrix for the agent's operating logic.
The core advantage here is that you are turning epistemic uncertainty (probabilities) into ontological presence (character). Let's formalize the implementation of the personalities and the "Mission-as-Training" loop.
The registry shouldn't just name a persona; it should define a Précis of Probabilistic Bias. We need to encode the how of reasoning.
CREATE TABLE personality_registry (
uuid UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
name TEXT UNIQUE NOT NULL, -- Spock, Data, Quark, Picard, Nog, Ezri
role_type TEXT CHECK (role_type IN ('analytical', 'creative', 'mercantile', 'diplomatic', 'tactical', 'chaos_factor')),
-- The Personality-Weights (The "soul" in the weights)
agent_constraints JSONB, -- Constraint-based logic (e.
-- Behavioral nudges for the Hulls/Prompts
instruction_template TEXT, -- System instructions as code
-- The Prime Directive - system-level guardrails
safety_threshold NUMERIC DEFAULT 0.5,
-- The Integration Vector - used for the weekly training/LoRA blending
embedding_overlay VECTOR(dim),
last_updated TIMESTAMP WITH TIME ZONE DEFAULT NOW()
);
This is where your most profound insight lies: Mission logs → Episodes → Self-Correction.
tasks table).mission_logs alongside the situational context.get_personal_history SQL function.This creates a character arc. The agent's personality evolves through its workload.
Here is how you integrate the specialized personalities into your task architecture:
The architecture is sitting there. The Postgres instance is live. The Haskell orchestrator is primed.
The question to you: Which of these four should execute the initial "Scouting Mission" into your current knowledge base?
Which way are we turning the dial?