Flatbot

The official chat interface for Flatbot-micro-4M, a tiny language model trained from scratch.

Flatbot is the official playground for Flatbot-micro-4M, a compact conversational language model built entirely from scratch using the Flatseek ecosystem.

This repository bundles the model, a modern ChatGPT-style web interface, and OpenAI-compatible APIs into a single application, making it easy to explore, deploy, and experiment with one of the smallest end-to-end language models available.

Unlike projects focused solely on inference, Flatbot showcases the complete lifecycle of an LLM—from tokenizer training and transformer training to GGUF export, streaming inference, and browser-based interaction.

Try it online: https://chat.flatseek.io
Download the model: https://huggingface.co/flatseek/flatbot-micro-4M

---

Flatbot-micro-4M

Flatbot-micro-4M is the flagship demonstration model of the Flatseek ecosystem.

Rather than competing with large foundation models, it demonstrates how a complete conversational AI model can be built from scratch using open-source tooling. The model is intentionally compact, making it suitable for education, experimentation, edge deployments, and understanding how modern LLMs work internally.

Experimental model: Flatbot is intended for learning and experimentation. Expect occasional hallucinations, factual inaccuracies, and inconsistent responses.

Model Overview

Property	Value
Parameters	~4.8 Million
Size	~19 MB
Format	GGUF (F16)
Architecture	Decoder-only Transformer
Context Length	512 tokens
Runtime	Flatrun
Training	From random initialization

Architecture

Component	Details
Position Encoding	RoPE
Normalization	RMSNorm
Feed Forward	SwiGLU
Attention	Grouped Query Attention (GQA)
Weight Tying	Yes

Training

The model was trained entirely from random initialization using FlatBuild.

Property	Value
Conversations	~2,500
Training Split	95%
Validation Split	5%
Dataset	General conversational dialogue

---

Chat Interface

Flatbot includes a lightweight browser-based interface for interacting with Flatbot-micro-4M.

Live Demo

https://chat.flatseek.io

Features

• ChatGPT-style interface
• Real-time streaming responses
• Bundled Flatbot-micro-4M model
• CPU-only inference
• No external inference server
• OpenAI-compatible API
• Anthropic-compatible API
• Deployable on Vercel

---

API

Flatbot exposes APIs compatible with popular LLM clients.

OpenAI

POST /v1/chat/completions

Anthropic

POST /v1/messages

Additional Endpoints

GET /v1/models
GET /healthz

Example

curl https://chat.flatseek.io/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model":"flatbot",
    "messages":[
      {
        "role":"user",
        "content":"Hello!"
      }
    ]
  }'

Streaming is enabled with:

{
  "stream": true
}

---

Running Locally

pip install -r requirements.txt

uvicorn api.index:app --reload --port 8000

Open:

http://localhost:8000

---

Environment Variables

Variable	Default	Description
MODEL_PATH	./model	Path to the GGUF model
CACHE_MB	256	KV cache size

---

Limitations

• CPU-only inference
• Cold-start model loading
• One active generation per instance
• Vercel Hobby plans have a 10-second execution limit

---

Part of the Flatseek Ecosystem

• Flatseek — Keyword Search
• Flatvec — Vector Search
• Flatask — RAG Runtime
• Flatlens — Data Visualization
• Flatbuild — LLM Training
• Flattune — LLM Fine-Tuning
• Flatrun — LLM Inference

---

Flatbot demonstrates that a complete conversational AI stack—from training a language model to serving it through a modern chat interface—can fit into a compact, fully open-source project.