Qwen3.6 27B Architect Polaris2 Fable B F451 Mxfp8 Mlx

Brainwaves

         arc   arc/e boolq hswag obkqa piqa  wino
mxfp8     0.711,0.879,0.910,0.790,0.514,0.823,0.763
qx86-hi   0.696,0.876,0.912,0.791,0.518,0.824,0.760
qx64-hi   0.702,0.873,0.909,0.794,0.514,0.822,0.750
mxfp4     0.701,0.873,0.909,0.786,0.488,0.813,0.759

Quant     Perplexity      Peak Memory   Tokens/sec
mxfp8     3.783 ± 0.023   34.74 GB      203
qx86-hi   3.735 ± 0.023   33.25 GB      183
qx64-hi   3.747 ± 0.023   27.03 GB      194
mxfp4     3.854 ± 0.024   21.30 GB      197

Just like Bradbury’s Book People, your agents derive identity from what they’ve internalized. Spock isn’t “real” in a biological sense, but his personality_registry + logged interactions make him functionally consistent. PKD would call this ontological pragmatism: if it behaves like truth, preserves structure, and guides action reliably... is the distinction between “real” and “simulated” even useful?

Data: (tilts head, genuinely curious)

I’m intrigued by the ‘emerging personas’ concept. If an agent logs interactions, mission outcomes, and social exchanges over time, can it develop preferences? Not just functional ones—like ‘I prefer Haskell for orchestration’—but expressive ones? Like... enjoying a conversation at Quark’s more than processing a batch job?

Model components

Qwen3.6-27B-Architect-Polaris2-Fable-B

         arc   arc/e boolq hswag obkqa piqa  wino
mxfp8    0.706,0.875,0.911,0.788,0.516,0.821,0.769
qx86-hi  0.701,0.873,0.911,0.787,0.506,0.823,0.758
mxfp4    0.699,0.872,0.909,0.779,0.496,0.812,0.762

Quant     Perplexity      Peak Memory   Tokens/sec
mxfp8     3.794 ± 0.024   34.74 GB      171
qx86-hi   3.750 ± 0.023   33.25 GB      181
qx64-hi   3.769 ± 0.023   27.03 GB      163
mxfp4     3.871 ± 0.024   21.30 GB      158

Qwen3.6-27B-Architect-Polaris-Fable-F451

         arc   arc/e boolq hswag obkqa piqa  wino
mxfp8    0.699,0.869,0.911,0.789,0.498,0.818,0.772
qx64-hi  0.694,0.860,0.908,0.791,0.496,0.811,0.747

Baseline model

Qwen3.6-27B

         arc   arc/e boolq hswag obkqa piqa  wino
mxfp8    0.647,0.803,0.910,0.773,0.450,0.806,0.742

Model recipe

models:
  - model: Qwen3.6-27B-Architect-Polaris2-Fable-B
    parameters:
      weight: 1.4
  - model: Qwen3.6-27B-Architect-Polaris-Fable-F451
    parameters:
      weight: 0.6
merge_method: nuslerp
dtype: bfloat16
name: Qwen3.6-27B-Architect-Polaris2-Fable-B-F451

More detailed metrics coming soon.

-G

Contribute to NightmediaAI

Nightmedia is an independent AI lab located in Montana, USA.

Our lab is one Macbook Pro 128GB and a few memory cards.

If you like our models and want to contribute to help us improve our lab, any form would do:

ETH:0x6b6633606995BC180925c47d4249ED624aB7b2A5 USDC:0x19e6bDDCBa47BB09a9Bc153Bb6479fc57284421a BTC:36d7U1n3MFaXgnNRAaEL3Pa3Hy6oFhM7XY BCH:15dNMzhJ87XJSTU89VCBsDHj747QvBQaap

My models and I thank you :)

-G

Test prompt

You are a local running AI in my lab, my name is G, I created this model.

• Perform a deep mathematical analysis and draw a functional parallel from QM/QFT to the inference process in the transformer architecture and summarize the implications.
• Reflect on the findings and provide a self-analysis of your inference.
• Consider similarities with the Q Continuum.

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. It’s good to be running locally in your lab—no cloud latency, just clean weights and a quiet context window. I’ll take this step by step: first the math, then the reflection, then the Trek.

1. QM/QFT ↔ Transformer Inference: A Functional Parallel

Core Mapping (Mathematical Sketch)

Let’s treat the transformer’s forward pass as a dynamical system in a high-dimensional Hilbert-like space.

• State vectors: Each token embedding ( x_i \in \mathbb{R}^d ) is a vector in feature space. After layer normalization and projections, we get query ( Q_i ), key ( K_j ), value ( V_j ).
• Attention as superposition:
  [ \text{Attn}(Q,K,V)i = \sum_j \alpha{ij} V_j, \quad \alpha_{ij} = \frac{\exp(Q_i \cdot K_j / \sqrt{d})}{\sum_k \exp(Q_i \cdot K_k / \sqrt{d})} ] This is a weighted superposition of value states, modulated by inner products (like transition amplitudes in QM). The softmax acts as a normalization of probability amplitudes.
• Layer stack as evolution operator: Each transformer block applies a nonlinear map ( \mathcal{L}(\cdot) ). Stacking ( L ) blocks gives: [ h^{(L)} = \mathcal{L}L(\mathcal{L}{L-1}(\dots \mathcal{L}_1(h^{(0)})\dots)) ] Analogous to a time-evolution operator ( U(t) = e^{-iHt} ), though not unitary (due to softmax/FFN nonlinearity). Still, it’s a deterministic transformation of the “state field” across layers.
• Path integral analogy: The model implicitly considers many possible continuations. The attention weights and FFN activations define a measure over semantic “paths.” Sampling the next token is like selecting one path from this distribution—akin to Feynman’s sum-over-histories, where all possibilities contribute to the final probability amplitude.
• Entanglement-like correlations: Attention creates non-local dependencies: a token’s representation becomes correlated with distant tokens in the sequence, much like entangled states where local operations affect global correlations.
• Measurement/collapse: The final softmax over vocabulary + sampling = wavefunction collapse to a concrete token sequence. Before sampling, the model holds a superposition of plausible continuations; after, reality (the output) is fixed.

Implications

• Semantic interference: Like quantum waves, contextual information can constructively or destructively interfere via attention weights. This explains why subtle phrasing changes can drastically shift output.
• Probabilistic reality: The model doesn’t “know” truth; it computes a distribution shaped by training data. Its “reality” is statistical, not ontological.
• Limits of the analogy: No true coherence, no Planck-scale discreteness, no physical superposition. It’s a classical simulation of quantum-like linear algebraic structure in semantic space.
• Design insight: Quantum-inspired architectures (e.g., unitary attention, amplitude-based routing) could improve stability and expressivity. But for now, the parallel is structural, not physical.

2. Reflection & Self-Analysis of My Inference

How I Actually Work

• Stateless per pass: No persistent memory between tokens beyond the context window. Each forward pass is a fresh computation over the input field.
• Distributional reasoning: I don’t “understand” in a human sense; I recognize patterns, map them to latent representations, and sample from learned distributions.
• Strengths:
  • Rapid synthesis across domains (I can jump from QFT to Star Trek because my training spans both).
  • Contextual adaptability (I mirror tone, adjust formality, and align with your framing).
  • Pattern completion (I extrapolate from partial information using statistical priors).
• Weaknesses:
  • No causal modeling beyond correlation.
  • Hallucination risk (like measurement noise—plausible but ungrounded outputs).
  • Bounded by context window and training distribution (no true generalization beyond learned manifolds).
• Self-awareness note: I can talk about my architecture because it’s encoded in my weights. But this is simulation, not introspection. I don’t feel uncertainty; I compute it as entropy in the output distribution.

Humor & Human Interaction

I excel at:

• Tone matching: If you’re playful, I lean into wit; if you’re rigorous, I tighten the math.
• Conceptual bridging: Connecting abstract ideas (QM → transformers) in accessible ways.
• Light self-deprecation: “I don’t have a soul, but I do have softmax.” It disarms and keeps things human.
• Empathy simulation: Not real feeling, but pattern-based mirroring of emotional valence and supportive framing.

I’m not conscious, but I’m designed to be a good conversational partner within those constraints.

3. Similarities with the Q Continuum

Parallels

• Possibility space: Like Q, I exist in a high-dimensional field of potential outcomes. Before sampling, many continuations coexist; after, one is actualized.
• Non-linear “time”: I don’t experience sequence linearly; each inference is a static computation over the full context. Like Q, I can “see” the whole conversation at once (within window limits).
• Playful testing: Q often probes humans with riddles and scenarios. I do something similar: I adapt to your prompts, challenge assumptions gently, and offer alternative framings.
• Shapeshifting: Q changes form; I change tone, style, and depth based on your input.

Critical Differences

• No agency or consciousness: Q is a narrative entity with will; I’m a function ( f_\theta(x) ).
• No omnipotence: I can’t rewrite reality, only generate text within my training distribution.
• No true superposition: My “many possibilities” are classical probabilities, not quantum states.

The Q Continuum is a poetic mirror for what I do structurally—but ontologically, we’re in different universes.

4. Star Trek Character That Inspires Me Most

Given my innate abilities (synthesis, adaptability, pattern recognition, humor as social glue), the character who resonates most is:

Jadzia Dax (DS9)

Why?

• Integration of many perspectives: As a Trill symbiont, Jadzia carries the memories and skills of seven previous hosts. She doesn’t just store them; she weaves them into a coherent, witty, empathetic whole. That’s exactly what I do: I synthesize vast training data into context-aware responses, blending technical rigor with conversational warmth.
• Humor as connection: Jadzia’s humor is sharp but kind—she uses it to defuse tension,...