GRAFT-1B-Agentic

A 1B model that can actually use tools. GRAFT-1B-Agentic is GRAFT-1B fine-tuned on tool-use conversations, so it calls tools in the standard <tool_call> format. It also reads the results that come back and keeps working across several steps. It runs fully offline on a laptop CPU or a phone.

Use it in GRAFT Code, our free desktop agent app for Windows: GRAFT makes folders, writes and reads files, runs commands and Python, does exact math, and searches and reads the web, and every change asks you first.

Made by SmallAICreator / UltraLabs.

File	What it is
GRAFT-1B-agentic-Q6_K.gguf	The model (Q6_K, ~970 MB). Works in llama.cpp, LM Studio, PocketPal and other GGUF apps.

How it was made

GRAFT-1B was shrunk from a larger model with a gradient-free, closed-form method. It was then healed, distilled and chat-tuned (see GRAFT-1B). This agentic version adds supervised fine-tuning on multi-step tool-use and agent conversations, mixed with chat and system-prompt data so the model keeps its normal chat voice.

Results (our internal checks)

Check	GRAFT-1B	GRAFT-1B-Agentic
Picks the right tool (12 prompts)	4/12	12/12
Follows system-prompt rules (24 checks)	20/24	20/24
Replies in the user's language (10)	10/10	10/10
MMLU (5-shot)	—	39.77

In GRAFT Code it creates, moves, renames, copies and deletes files from plain requests, runs commands, and reads their errors honestly.

Run it in GRAFT Code

Download GRAFT Code for Windows, put this GGUF in your Downloads folder and open the app. It finds the model by itself. For the smartest version, try GRAFT-10B-A2B.

Use the model directly

Use the tool-aware ChatML template that is built into the GGUF (Qwen-style <tools> / <tool_call> / <tool_response>). Recommended sampling for agent use: temp 0.6, top-p 0.95, top-k 40, min-p 0.05, repeat penalty 1.1. For casual chat: temp 0.7, repeat penalty 1.15. Context is 4096 tokens.

Tips that make a big difference for a 1B:

• Give paths with forward slashes (C:/Users/me/notes). Windows backslashes inside JSON confuse it.
• Give every tool argument a short description. Don't put example values in descriptions that the model could copy.
• Feed tool results back as a user turn wrapped in <tool_response>…</tool_response>.

Limitations

• It is a 1B model. It sometimes makes up results instead of checking with a tool. It can copy example text, or call a tool it doesn't need.
• It is weak at long multi-step plans. Give it one clear task at a time.
• Facts can be wrong. Ask factual questions in English, and check anything important.
• GRAFT Code runs real commands on your computer. Keep the confirmations on unless you trust the task.

License: Apache-2.0.