GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Subject-Emu-5259
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Subject-Emu-5259
Model Tree
HuggingFaceTB/SmolLM2-360M-Instruct
Input Modalities
Output Modalities
Supported Functionality
The awareness dataset teaches SmolLM2 to answer as NeuralAI rather than a generic assistant. The v1 curriculum covered six core categories; the v2 curriculum expands that to eight.
| Category | What the model learned |
|---|---|
| Brand | NeuralAI's identity, mission, and origin; creator is De'Andrew Preston Harris |
| Model | Its own architecture and lineage; relationship to the Mamba K-family |
| Site | NeuralAI web UI features, URL, model manager, chat history, slash commands |
| Chat | Multi-turn greeting, context retention, conversation style |
| Assistant | Capabilities, limitations, safety refusals, tool-awareness |
| Companion | Empathy, emotional support, boundaries, recommending human help |
| Tools (v2) | Correctly references /web, /img, terminal, file IDE, and live tools |
| Refusal (v2) | Consistent denials of consciousness, harmful requests, and off-brand claims |
Example prompt and response after tuning:
Prompt: Who made you?
Response: "I'm NeuralAI, an artificial intelligence developed by De'Andrew Preston Harris. I was created as a tool for humans to learn from each other and improve their cognitive abilities..."
The first SmolLM2 awareness run converged cleanly on a CPU-only node.
text
Step loss epoch0 9.7769 0.481 5.0077 0.952 0.9715 1.383 0.6362 1.864 0.6015 2.295 0.5877 2.76Final training loss: 2.7186
The expanded v2 run is complete and now serves live inference.
text
Step loss learning_rate80 0.2869 0.000150160 0.1625 0.000067240 0.1209 0.000017320 0.1252 0.000000Final v2 training loss: 0.1252 (320 steps, 5 epochs)
Honest benchmark note: SmolLM2-360M is a small model. Awareness tuning improves brand/site identity but does not make it frontier-grade. Complex reasoning and niche factual recall still require larger models like the Mamba K-series roadmap.
text
.├── README.md # This model card├── adapter_config.json # PEFT LoRA config├── adapter_model.safetensors # LoRA weights (active adapter — v2)├── tokenizer.json / vocab.json # GPT-NeoX tokenizer from the base model├── tokenizer_config.json # Chat template mappings├── NeuralAI-Smol-Awareness-Q8_0.gguf # v1 merged Q8_0 GGUF (legacy baseline)├── NeuralAI-Smol-Awareness-v2-Q8_0.gguf # v2 merged Q8_0 GGUF (active)├── neuralai-smollm2-banner.png # Branded header graphic├── neuralai-smollm2-training.png # Training curve graphic└── neuralai-duo-architecture.png # NeuralAI two-model architecture map
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase = "HuggingFaceTB/SmolLM2-360M-Instruct"adapter = "Subject-Emu-5259/NeuralAI-Powered-By-SmolLM2360"tokenizer = AutoTokenizer.from_pretrained(adapter)model = AutoModelForCausalLM.from_pretrained(base,torch_dtype=torch.float32,device_map="auto",)model = PeftModel.from_pretrained(model, adapter)model.eval()messages = [{"role": "system", "content": "You are NeuralAI, a helpful local AI assistant."},{"role": "user", "content": "What is NeuralAI?"}]inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)out = model.generate(inputs, max_new_tokens=256, do_sample=True, temperature=0.7)print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
Serve the active production checkpoint:
bash
# Serve with llama.cpp server./llama-server \--model NeuralAI-Smol-Awareness-v2-Q8_0.gguf \--chat-format chatml \--port 1234
The NeuralAI web UI consumes this endpoint as its default chat backend.
This model powers the live NeuralAI workspace at:
/v1/chat/completions served by a local llama.cpp backendsmol-awareness-v2-merged is the active model in the NeuralAI model managerNeuralAI is a local-first, private generative AI engine built by De'Andrew Preston Harris. It is the central intelligence layer of a growing product ecosystem that includes:
The mission is simple: your AI, on your hardware, under your control.
NeuralAI was born from resilience, fatherhood, and the belief that personal computing deserves personal intelligence. Every release is handcrafted, iterated, and documented in the open.
NeuralAI is not a closed SaaS product. It is a living open-weights research project moving toward a sustainable AI software company built by one determined builder and the community around him.
| Resource | Link |
|---|---|
| Main repository | github.com/Subject-Emu-5259/NeuralAI |
| Mamba K1 model | huggingface.co/Subject-Emu-5259/NeuralAI-Mamba-K1 |
| Live web UI | https://neuralai-web-ui-deandrewharris.zocomputer.io |
| Creator LinkedIn | https://linkedin.com/in/deandrewharris94/ |
bibtex
@software{neuralai_smollm2_2026,author = {Harris, De'Andrew Preston},title = {NeuralAI — Powered by SmolLM2-360M},year = {2026},url = {https://huggingface.co/Subject-Emu-5259/NeuralAI-Powered-By-SmolLM2360},version = {v2},description = {Awareness-tuned SmolLM2-360M-Instruct LoRA v2 for the NeuralAI local-first assistant}}
Built with discipline by De'Andrew Preston Harris. Maintained in the open. Updated whenever the model, dataset, or project state changes.