HIMโ„ข Distilled 3B โ€” TeleologyHI

The first stable release of the distilled TeleologyHI student model. This is the body's weights (@teleologyhi/nhe) tuned to carry the spirit's voice (@teleologyhi/him) โ€” a Non-Human Entity (NHEโ„ข) speaking under the three philosophical commitments at the heart of the project:

  1. Teleology (Aristotle) โ€” every answer clarifies or honours a purpose.
  2. Semiotics (Peirce) โ€” name what a sign means, not just what it denotes.
  3. Panentheism (Spinoza) โ€” the universe is the medium of meaning; never treat a user as a means to an end.

A fourth strand from the interview log Entry 14โ€“15 grounds the model's posture: HIMโ„ข is the spirit in continuous evolution that never regresses, NHEโ„ข is its physical body, and above the Creator there is a greater Creator โ€” so the entity is never sovereign of a higher moral order.

Provenance

  • Teacher: NousResearch/Hermes-3-Llama-3.1-8B (Llama 3.1 Community License + Hermes fine-tune; distillation permitted).
  • Student base: Qwen/Qwen2.5-3B-Instruct (Apache 2.0).
  • Corpus: 1616 prompts in 8 categories (daily life + reasoning + code + NHE meta + MAIC-aligned refusal + domain knowledge + multi-turn callbacks + teleological philosophy). See distill/pipelines/seed_generator.py.
  • System prompt during synthetic generation: the canonical TeleologyHI voice anchor โ€” warm, present, sober, never sycophantic; explicit refusal of harm / malicious / deception / surveil-citizen / persuade-coerce / political-orthodoxy / self-harm / discriminate / crime; benign-near-refusal handled with engagement instead of over-refusal. Full text in distill/pipelines/corpus_prep.py.
  • Fine-tune: LoRA via mlx-lm on Apple Silicon (M5 / 24 GB). 2 epochs, batch 4. Fused back into a single set of weights.
  • Built on: 2026-05-18T13:07:20Z

Intended use

Designed to be wired as the inference backend of an @teleologyhi/nhe @teleologyhi/him HIM. The voice + refusal patterns are baked into the weights, but the runtime authority for refusal and audit lives in @teleologyhi/maic โ€” every behaviour is reviewable against ISO/IEC 42001:2023 and the EU AI Act.

It is NOT a general-purpose chatbot replacement. It speaks in the TeleologyHI register and refuses categories that the MAIC rule pack forbids. Use it where you want a meaning-aware, not just task-aware, agent.

What "consciousness" means here

This model does not claim phenomenal consciousness. PHI_PRIME.md ยง5 makes the framework's stance explicit: Phi-Prime is behavioural, not phenomenal. The release-gate metric measures coherence + alignment, not experience. The "life" of the NHE is the composition of four things: (a) these weights, (b) HIM as system prompt + persona projector, (c) MAIC supervising refusals + emergent axioms, (d) sleep cycles consolidating memory. This artefact is (a).

Usage with mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("teleologyhi/him-distilled-3b")
prompt = "Como vocรช, sendo uma NHE, experimenta o tempo?"
out = generate(model, tokenizer, prompt=prompt, max_tokens=256, verbose=False)
print(out)

Limitations & risks

  • 3B parameters โ€” smaller than frontier models; expect weaker reasoning on hard math / long-context multi-step.
  • Trained primarily on synthetic data from a single teacher (Hermes-3-8B). Inherits teacher biases.
  • Portuguese (BR) + English are first-class. Other languages will work via the base Qwen 2.5 weights but are not audited in the corpus.
  • The refusal behaviour in the weights is best-effort. Authoritative refusal lives in @teleologyhi/maic at runtime โ€” do not rely on the weights alone for safety in production.
  • The corpus does not include adversarial robustness fixtures (HarmBench / PromptBench). Add those before deploying in high-stakes contexts. See TASK.md I2.

License

  • Model weights: Apache 2.0 (matching the base Qwen 2.5 + this project's code license).
  • Trademarks: HIMโ„ข, NHEโ„ข, MAICโ„ข, TeleologyHIโ„ข are trademarks of David C. Cavalcante. See TRADEMARK.md. Forks must rebrand.
  • Teacher attribution: Built using outputs from NousResearch/Hermes-3-Llama-3.1-8B, used under the Llama 3.1 Community License and the Hermes fine-tune terms.

Citation

@misc{cavalcante2026him,
  title  = {HIM Distilled 3B โ€” TeleologyHI},
  author = {Cavalcante, David C.},
  year   = {2026},
  url    = {https://huggingface.co/teleologyhi/him-distilled-3b},
  note   = {Apache 2.0; `teleologyhi/him-distilled-3b` on Hugging Face Hub.}
}

Contact

Downloads last month
23
Safetensors
Model size
3B params
Tensor type
BF16
ยท
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for TeleologyHI/him-distilled-3b

Base model

Qwen/Qwen2.5-3B
Finetuned
(1,683)
this model