| --- |
| language: |
| - en |
| license: apache-2.0 |
| library_name: transformers |
| inference: true |
| pipeline_tag: text-generation |
| tags: |
| - text-generation |
| - business-intelligence |
| - market-analysis |
| - web-research |
| - agent |
| - tool-calling |
| - business-consultancy |
| - autonomous-agent |
| - competitive-intelligence |
| - trend-forecasting |
| - chatpbc |
| - luwa |
| --- |
| |
| # luwa-01 |
|
|
| ## Experimental Open Source retrieval-assisted business research AI. |
|
|
| luwa-01 is experimental, open source retrieval-assisted business research AI model built by ChatPBC. It is designed for one purpose: **delivering expert-level market analysis, business strategy, and competitive intelligence** — powered by real-time web research through an integrated agent framework. |
|
|
| Version: v5 · Built by ChatPBC · Developed by Mik Tse · [Documentation](https://chatpbc.xo.je/luwa) · [Live Playground](https://chatpbc.xo.je/luwa) · [Hugging Face](https://huggingface.co/chatpbc1/luwa-01) |
|
|
| --- |
|
|
| ## Why luwa-01 Exists |
|
|
| In 2026, businesses face an unprecedented information crisis. The speed of market change has accelerated to the point where traditional research — hiring consultants, commissioning reports, waiting weeks for analysis — is simply too slow. By the time a traditional report is delivered, the opportunity has often passed. |
|
|
| The critical problems businesses face today: |
|
|
| | Problem | Impact | |
| |---------|--------| |
| | **Information overload** | 2.5 quintillion bytes generated daily; impossible to process manually | |
| | **Speed pressure** | Competitors make decisions in minutes, not months | |
| | **Cost constraints** | McKinsey charges $50K+ per engagement; Bloomberg Terminal costs $24K/year | |
| | **Analysis paralysis** | Too much data, too little actionable insight | |
| | **Talent shortage** | Business intelligence analysts are expensive and scarce | |
|
|
| luwa-01 solves all five problems simultaneously. It provides business-focused analysis, near-instant responses on CPU, and zero API costs — all with real-time web access through its agent tool-calling framework. |
|
|
| --- |
|
|
| ## How Businesses Use luwa-01 |
|
|
| luwa-01 is deployed across multiple business functions: |
|
|
| **Market Entry Analysis** — Before entering a new market, businesses query luwa-01 for size, growth, competition, and barriers to entry. It synthesizes real-time data into a decision-ready report in seconds. |
|
|
| **Competitive Intelligence** — luwa-01 monitors competitors, analyzes their positioning, pricing, and product launches, and delivers strategic recommendations on how to differentiate. |
|
|
| **Trend Forecasting** — For strategic planning, luwa-01 forecasts industry trends 1-5 years out, identifying emerging opportunities before competitors do. |
|
|
| **Business Validation** — Entrepreneurs and corporate innovation teams use luwa-01 to validate business ideas, assess market fit, and score opportunities against key criteria. |
|
|
| **Investment Research** — Analysts use luwa-01 to quickly assess sectors, identify investment opportunities, and generate preliminary due diligence. |
|
|
| --- |
|
|
| ## The Critical Problem luwa-01 Solves in 2026 |
|
|
| The main problem in 2026 is not a lack of information — it's the inability to **act on information fast enough**. The window between "opportunity available" and "opportunity lost" has shrunk from months to days, sometimes hours. |
|
|
| luwa-01 compresses this decision loop to **under a second** (on CPU). A business that queries luwa-01 gets: |
|
|
| 1. **Real-time market data** (via web research tools) |
| 2. **Professional analysis** (structured like a consulting report) |
| 3. **Actionable recommendations** (specific next steps, not vague advice) |
| 4. **Risk assessment** (what could go wrong and how to mitigate it) |
|
|
| This is the difference between reacting to the market and **shaping** it. |
|
|
| --- |
|
|
| ## luwa-01 vs General-Purpose Models |
|
|
| luwa-01 is not designed to replace GPT-4o or Claude. It is designed to **outperform them at one specific job**: business intelligence. |
|
|
| | Dimension | luwa-01 | GPT-4o | Claude 3.5 | Gemini Pro | Mistral 7B | |
| |-----------|---------|--------|------------|------------|------------| |
| | **Business specialization** | 100% dedicated | General-purpose | General-purpose | General-purpose | General-purpose | |
| | **Response latency** | ~50ms | 3-5 seconds | 2-4 seconds | 3-5 seconds | 500ms | |
| | **Cost per 1M tokens** | $0.02 | $10.00 | $15.00 | $3.50 | $0.25 | |
| | **Self-hostable** | Yes (CPU or GPU) | No | No | No | Yes | |
| | **Open weights** | Yes (Apache 2.0) | No | No | No | Yes | |
| | **Requires internet** | No (works offline) | Yes | Yes | Yes | No | |
| | **Data privacy** | 100% local | Cloud-dependent | Cloud-dependent | Cloud-dependent | Local | |
| | **Customizable** | Fully | Limited | Limited | No | Fully | |
| | **Runs on CPU** | Yes (4GB RAM) | No | No | No | Marginal | |
| | **Model size** | 942 MB | ~30 GB | ~25 GB | ~20 GB | 14 GB | |
| | **Tool-calling for business** | Native | API-dependent | API-dependent | API-dependent | None | |
| | **Agent-first design** | Yes | No | No | No | No | |
| | **Native research decision** | Yes (model decides) | Prompt-dependent | Prompt-dependent | Prompt-dependent | None | |
| | **Mode signals** | [RESEARCH]/[CONVERSATIONAL] | None | None | None | None | |
| | **Research synthesis** | Yes (structured analysis) | Yes | Yes | Yes | No | |
|
|
| ### Why Businesses Should Choose luwa-01 |
|
|
| **1. Speed.** Real-time decisions require real-time answers. luwa-01 responds in under a second on CPU — fast enough for interactive use. No GPU required. |
|
|
| **2. Cost.** luwa-01 runs on free CPU hardware. No API costs, no per-token billing. A business can run unlimited queries for $0 — just electricity. |
|
|
| **3. Privacy — your competitive intelligence never leaves your infrastructure.** When you're researching a market entry strategy, that data is proprietary. luwa-01 runs entirely on your hardware. |
|
|
| **4. Control — full ownership.** Apache 2.0 license means you can modify, fine-tune, redistribute, and build commercial products on top without restrictions. |
|
|
| **5. Specialization.** luwa-01 is trained specifically for business intelligence tasks. Its parameters are dedicated to market analysis, strategy, and business consultancy — not general-purpose chat. |
|
|
| **6. Agent-first design.** luwa-01 was built from the ground up as an agent, not a chatbot. Its tool-calling interface is native, not bolted on. |
|
|
| **7. Native decision-making.** The model itself decides when to research the web vs. when to answer directly from its knowledge. It outputs [RESEARCH] or [CONVERSATIONAL] signals that the agent framework interprets — no external routing needed. |
|
|
| --- |
|
|
| ## Model Architecture |
|
|
| luwa-01 is a transformer-based language model optimized for business intelligence tasks. |
|
|
| | Parameter | Value | |
| |-----------|-------| |
| | Parameters | ~500M | |
| | Hidden size | 896 | |
| | Attention heads | 14 (4096 dim, GQA with 2 KV heads) | |
| | Layers | 24 | |
| | Vocabulary | 152,064 tokens | |
| | Max context | 32,768 tokens | |
| | Weight format | safetensors (942 MB) | |
| | License | Apache 2.0 | |
| | Latency | ~1-5s on CPU (500M params, no GPU needed) | |
| | Cost per query | $0 (self-hosted) | |
|
|
| --- |
|
|
| ## Capabilities |
|
|
| luwa-01 is not a general-purpose chatbot. It is a **specialized agent** with the following capabilities: |
|
|
| | Capability | Description | |
| |-----------|-------------| |
| | **Market Analysis** | Size, growth, segmentation, competitive landscape | |
| | **Tool-Calling** | Outputs structured tool signals for web search, scraping, data retrieval | |
| | **Multi-Tool Chains** | Designed to chain multiple research tools (currently web_search only) |
| |
| --- |
| |
| ## Agent Framework |
| |
| luwa-01 has **native agent decision-making** built into the model. The model itself outputs decision signals that determine its behavior: |
| |
| - **[RESEARCH]** — The model recognizes the query requires current web data. The agent performs real-time web search with citations. |
| - **[CONVERSATIONAL]** — The model answers directly from its training knowledge (identity, greetings, general advice). |
| - **[TOOL_CALL]** — When in research mode, the model specifies its search query: |
| |
| ``` |
| [TOOL_CALL]{"tool": "web_search", "parameters": {"query": "AI market size 2026"}}[/TOOL_CALL] |
| ``` |
| |
| This is a NATIVE model capability — the model was trained to make this decision autonomously. No external routing or classification is needed. |
| |
| Available tools: |
| - `web_search` — Real-time web search via DuckDuckGo HTML scraping (currently implemented) |
| |
| Additional tools (`web_scrape`, `market_data`, `news_analysis`, `competitor_analysis`, `trend_forecast`) are designed but not yet implemented. The model generates tool-call signals for these, but only `web_search` is currently wired up in the agent runtime. |
| |
| The agent loop: **Model Decides → Parse Signal → Execute → Analyze → Synthesize → Deliver** |
| |
| The agent is also available as a live demo — see the [Try luwa-01](https://chatpbc.xo.je/luwa) page in the documentation. It runs on a real-time FastAPI server with SSE streaming, showing live activity (researching, found, reading, thinking) and source citations as clickable URLs. |
| |
| --- |
| |
| ## Quick Start |
| |
| ### Python (with transformers) |
| |
| ```python |
| from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
| model = AutoModelForCausalLM.from_pretrained( |
| "chatpbc1/luwa-01", |
| trust_remote_code=True, |
| device_map="auto" |
| ) |
| tokenizer = AutoTokenizer.from_pretrained("chatpbc1/luwa-01") |
| |
| prompt = "Analyze the AI market in 2026: size, growth, key players, and opportunities" |
| messages = [{"role": "user", "content": prompt}] |
| |
| text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) |
| inputs = tokenizer(text, return_tensors="pt").to(model.device) |
| outputs = model.generate(**inputs, max_new_tokens=512) |
| response = tokenizer.decode(outputs[0], skip_special_tokens=True) |
| print(response) |
| ``` |
| |
| ### Via API (Agent Runtime) |
| |
| ```bash |
| curl -X POST "https://your-url/agent" \ |
| -H "Content-Type: application/json" \ |
| -d '{"prompt": "Analyze the e-commerce market in Southeast Asia", "session_id": "abc123", "max_tokens": 512}' |
| ``` |
| |
| The agent returns SSE events for each activity step, followed by the final response with citations: |
| |
| ```json |
| {"type": "thinking", "message": "luwa-01 is thinking..."} |
| {"type": "researching", "message": "Researching the web..."} |
| {"type": "found", "message": "Found 5 relevant sources"} |
| {"type": "reading", "message": "Analyzing sources..."} |
| {"type": "thinking", "message": "luwa-01 is analyzing..."} |
| {"response": "...", "citations": [{"index": 1, "title": "...", "url": "..."}], "mode": "research"} |
| ``` |
| |
| The `mode` field in the final response indicates the model's decision: |
| - `"mode": "research"` — Model chose to research the web |
| - `"mode": "conversational"` — Model answered directly from knowledge |
| |
| ### Agent REST API Endpoints |
| |
| | Endpoint | Method | Description | |
| |----------|--------|-------------| |
| | `/health` | GET | Health check with model status | |
| | `/agent` | POST | Full agent flow with SSE streaming | |
| | `/chat` | POST | Simple chat (no web research) | |
| |
| --- |
| |
| ## Production Readiness Checklist |
| |
| - **Model format**: safetensors (safe, fast loading) |
| - **Tokenizer**: Full 152K vocab tokenizer with chat template |
| - **Generation config**: Optimized temperature (0.7), top_p (0.9), repetition penalty (1.1), no_repeat_ngram (3) |
| - **System prompt**: Business intelligence specialist persona |
| - **Agent config**: 6 tools, structured agent loop, confidence thresholding |
| - **Inference flag**: `inference: true` — ready for HF Inference API |
| - **License**: Apache 2.0 — commercial use permitted |
| - **Size**: 942 MB — fits on any GPU, even CPU (4GB RAM) |
| - **No external dependencies**: All files self-contained in the repo |
| |
| --- |
| |
| ## File Structure |
| |
| ``` |
| luwa-01/ |
| ├── model.safetensors # Model weights (942 MB) |
| ├── config.json # Architecture configuration |
| ├── generation_config.json # Optimized generation parameters |
| ├── tokenizer.json # Fast tokenizer (152K vocab) |
| ├── tokenizer_config.json # Tokenizer settings |
| ├── chat_template.jinja # Chat formatting template |
| ├── README.md # This documentation |
| ├── system_prompt.txt # Business intelligence system prompt |
| ├── agent_config.json # Agent tools and configuration |
| ``` |
| |
| --- |
| |
| ## Limitations |
| |
| - **Knowledge cutoff**: Training data reflects knowledge up to the base model's cutoff |
| - **Web access requires agent wrapper**: The model itself doesn't browse the web — it signals the agent to do so |
| - **500M size**: Not as capable as 70B+ models on complex multi-hop reasoning, but dramatically faster and cheaper |
| - **English primary**: Optimized for English business content |
| - **Synthesis depth**: The model's synthesis is good but not perfect — it may occasionally oversimplify or repeat patterns for very complex multi-domain queries |
| |
| --- |
| |
| ## Training Data |
| |
| The model was developed using **7 specialized synthetic datasets** totaling **~1,000 training samples** across 5 training iterations (v1-v5). The current model is v5. |
|
|
| | # | Dataset | Samples | Purpose | |
| |---|---------|---------|---------| |
| | 1 | Identity (Rigid) | ~80 | luwa-01 / Mik Tse / ChatPBC family | |
| | 2 | Conversational | ~35 | Greetings, Q&A, general business advice | |
| | 3 | Web Research | ~15 | [RESEARCH] signal + tool-call generation | |
| | 4 | Mixed (Research + Cite) | ~10 | Research with inline source citations | |
| | 5 | Market Analysis | ~230 | Market sizing, growth analysis, segmentation | |
| | 6 | Business Consultancy | ~360 | Strategic advice, recommendations, frameworks | |
| | 7 | Tool-Calling / Web Research | ~290 | Structured tool-call generation, search queries | |
|
|
| ### Native Decision Signals (v5) |
|
|
| The v5 model was specifically trained to output decision signals that enable autonomous agent behavior, plus **synthesis capability**: |
|
|
| | Signal | Meaning | Agent Behavior | |
| |--------|---------|----------------| |
| | `[RESEARCH]` | Query requires real-time data | Perform web search, cite sources, then synthesize | |
| | `[CONVERSATIONAL]` | Can answer from knowledge | Respond directly, no search | |
| | `[TOOL_CALL]{...}` | Specific search query | Execute the specified tool | |
|
|
| ### v5 Improvements |
|
|
| - **Synthesis**: After web research, the model synthesizes the data into structured analysis with key findings, recommendations, and inline citations |
| - **Rigid mode distinction**: The model decisively chooses between research and conversational modes |
| - **EOS token removal**: Responses no longer contain the model's internal special EOS marker (model's raw output includes an extra token that is stripped by the runtime) |
| - **Diverse responses**: The model generates varied, context-aware responses instead of repeating the same pattern |
|
|
| --- |
|
|
| ## License |
|
|
| Apache License 2.0 — Free for commercial use, modification, and distribution. |
|
|
| --- |
|
|
| ## Built by ChatPBC |
|
|
| luwa-01 is part of the ChatPBC ecosystem of specialized business AI models. |
|
|
| **Repository**: [huggingface.co/chatpbc1/luwa-01](https://huggingface.co/chatpbc1/luwa-01) |
|
|
