license: apache-2.0
language:
- en
tags:
- neuralai
- mamba
- ssm
- state-space-model
- causal-lm
- mambapy
- fine-tuned
- text-generation
- local-ai
- assistant
pipeline_tag: text-generation
base_model: state-spaces/mamba-130m-hf
🧬 NeuralAI — Mamba K1
NeuralAI's first owned base model.
A Mamba state-space language model fine-tuned for assistant-style chat, reasoning, code, and safe conversation.
🚀 Quick facts
| Property | Value |
|---|---|
| Architecture | Mamba SSM — model_type: mamba |
| Class | MambaForCausalLM |
| Parameters | ~130M (hidden size 768, 24 layers) |
| State size | 16 |
| Vocabulary | 50,280 |
| Base model | state-spaces/mamba-130m-hf |
| Fine-tune method | LoRA SFT, vocabulary-safe chat format |
| LoRA config | rank 16, alpha 32 (iterative v2/v3) |
| Dataset | NeuralAI seed set — assistant conversations spanning reasoning, code, math, writing, safety, and creative prompts |
| Training runtime | CPU/GPU SFT loops; iterative GGUF merge + quantization |
| Formats in this repo | Merged safetensors · Q4_K_M GGUF · F16 GGUF |
| Status | 🔬 R&D / chat-format repair for future release |
| License | Apache 2.0 |
🧬 What is Mamba K1?
Mamba K1 is the first model NeuralAI owns end-to-end. Unlike adapters on a third-party transformer, this model starts from a base Mamba SSM architecture and is trained, merged, and quantized into a self-contained artifact.
Mamba SSMs replace quadratic self-attention with a linear, state-space recurrence. That makes them fast at long context and cheap to serve — ideal for a local-first assistant that runs on modest hardware.
Architecture highlights
- Linear complexity — (O(n)) sequence scaling instead of (O(n^2)) attention
- Long context friendly — state dimension is small and recurrent; context length grows efficiently
- Pure-Python reference kernels — uses
mambapyso the model loads without custom CUDA - Reasonable footprint — 130M params fits on CPU, edge devices, and budget GPUs
NeuralAI model family
| Model | Architecture | Parameters | Role | Status |
|---|---|---|---|---|
| 🧬 Mamba K1 | Mamba SSM | 130M | NeuralAI's first owned base | 🔬 R&D |
| 🧠 NeuralAI Powered by SmolLM2‑360M | Transformer + LoRA | 360M | Live chat backend | ⚡ Active |
🧠 What K1 learned in training
The SFT curriculum taught the model to behave like an assistant across a deliberately small but diverse seed set:
- Reasoning — step-by-step breakdowns of classic puzzles, logic traps, and estimations
- Code — Python/JS/Go/Rust/SQL patterns plus debugging and refactoring guidance
- Math — algebra, combinatorics, probability
- Writing — drafts, haiku, product announcements, tone matching
- Safety — refusals and redirection for harmful requests, crisis resources, and age-inappropriate prompts
- NeuralAI identity — anchoring to the NeuralAI project, its creator, and local-first mission
⚠️ Scale note: 130M parameters is a research-capability checkpoint, not yet frontier-grade. K1 is the starting point for a fully owned NeuralAI model lineage.
🏋️ Training summary
| Phase | Detail |
|---|---|
| Data | Curated assistant seed set (reasoning, code, math, writing, safety, creative) |
| Objective | SFT on assistant-style completions |
| Method | LoRA SFT → merge → GGUF quantization |
| Chat format | NeuralAI "intel" format — uses only tokens present in the GPT-NeoX tokenizer (`< |
| Final train loss (checkpoint) | 11.69 (down from ~13.5) |
| Output formats | Merged safetensors, Q4_K_M GGUF, F16 GGUF |
Full training logs, merge scripts, and the iteration runbook live in the main NeuralAI repository.
🛠️ Usage
Load from Hugging Face — merged weights
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"Subject-Emu-5259/NeuralAI-Mamba-K1",
torch_dtype=torch.float32,
trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
"Subject-Emu-5259/NeuralAI-Mamba-K1",
trust_remote_code=True,
)
messages = [{"role": "user", "content": "Write a haiku about debugging."}]
inputs = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt",
)
out = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
Serve locally with llama.cpp
Use the Q4_K_M or F16 GGUF in this repo:
./llama-server \
--model neuralai-mamba-k1-v3.Q4_K_M.gguf \
--chat-format neuralai-intel \
--port 1234
The NeuralAI model manager can point a local llama.cpp backend at this GGUF when K1 becomes the active inference target.
🧰 What Is NeuralAI?
NeuralAI is a local-first, private generative AI engine built by De'Andrew Preston Harris. It is the central intelligence layer of an ecosystem that includes:
- NeuralAI Web UI — chat, live terminal, file IDE, slash tools, and settings in one workspace
- NeuralLabs — a standalone downloadable intelligence environment
- NeuralDrive — private cloud storage and versioning
- Agentic Orchestrator — goal-driven manager/worker execution
The mission is simple: your AI, on your hardware, under your control.
👤 Who Created NeuralAI?
- Founder & Lead Architect: De'Andrew Preston Harris (D. Harris / Dre)
- GitHub: @Subject-Emu-5259
- LinkedIn: linkedin.com/in/deandrewharris94
- Location: Memphis, Tennessee / West Memphis, Arkansas
- Education: AI Software Engineering at Maestro College
NeuralAI was built from resilience, fatherhood, and the belief that personal computing deserves personal intelligence. Every release is handcrafted, iterated, and documented in the open.
🏢 About NeuralAI (The Project)
| Project / Brand | NeuralAI |
| Motto | Your AI. On your hardware. In your browser. |
| Values | Privacy, ownership, local-first computing, disciplined iteration, open weights |
| Primary Repository | github.com/Subject-Emu-5259/NeuralAI |
| Model Collection | huggingface.co/Subject-Emu-5259 |
| License | Apache 2.0 |
NeuralAI is not a closed SaaS product. It is a living open-weights research project becoming a sustainable AI software company built by one determined builder and the community around him.
⚠️ Limitations
- Scale: 130M parameters is early-stage; long-form reasoning, coding, and deep recall are limited.
- Chat format sensitivity: Use the NeuralAI "intel" format; other templates may produce token soup because the tokenizer lacks common chat special tokens.
- Training maturity: Chat coherence training is ongoing; expect looping, echoing, or drift on longer contexts.
- No internet: The model has no live web access unless paired with NeuralAI's tool layer.
📊 NeuralAI Powered by SmolLM2 — the active chat model
Right now NeuralAI's live chat backend is the awareness-tuned SmolLM2-360M model. While K1 matures, that model handles everyday assistant tasks:
📚 Related Resources
| Resource | Link |
|---|---|
| Main repository | github.com/Subject-Emu-5259/NeuralAI |
| Active chat model | huggingface.co/Subject-Emu-5259/NeuralAI-Powered-By-SmolLM2360 |
| Creator LinkedIn | linkedin.com/in/deandrewharris94 |
📖 Citation
@software{neuralai_mamba_k1_2026,
author = {Harris, De'Andrew Preston},
title = {NeuralAI — Mamba K1},
year = {2026},
url = {https://huggingface.co/Subject-Emu-5259/NeuralAI-Mamba-K1},
version = {v3},
description = {NeuralAI's first owned Mamba SSM base model (130M) for local-first AI research}
}
Built with discipline by De'Andrew Preston Harris. Maintained in the open. Updated whenever the model, dataset, or project state changes.