license: mit
library_name: pytorch
tags:
- tool-calling
- agent
- tiny-llm
- byte-level
- on-device
- from-scratch
pipeline_tag: text-generation
ultra-tiny-1m — LocalAgent (0.98M params)
A from-scratch, byte-level tool-calling agent model from LocalAgent. Pure PyTorch, 0.98M params, trained on CPU. It pairs a tiny decoder (GQA + RoPE + SwiGLU + depth-recurrence) with a dual head (tool-selection classifier + pointer/copy argument head) and prompt-grounded constrained decoding for reliable tool calls across 21 tools (general assistant, the Claude Code / Codex coding surface, and computer-use / productivity tools), including parallel two-call turns.
Architecture
- vocab 256 (byte-level), d_model 192, layers 2 x6 loops, heads 6/2 (GQA), ffn 640
- factorized embeddings: True
Files
config.json—ModelConfigmodel.safetensors/pytorch_model.bin— decoder weightsagent_heads.bin— trained tool-selection + pointer heads (optional)
What it can do (use cases)
One byte-level model that turns a natural-language turn into a grounded tool call — across an assistant, a coding agent, computer-use/productivity apps, and parallel two-call turns:
| you say | it calls |
|---|---|
| "What's the weather in Cusco?" | get_weather(city="Cusco") |
| "What is 19 * 19 * 5?" | calculator(expression="19*19*5") |
| "Open the file bin/run.sh." | read_file(path="bin/run.sh") |
| "Grep for 'TODO'." | grep_search(pattern="TODO") |
| "Run the tests." | run_tests() |
| "Commit with message 'fix bug'." | git_commit(message="fix bug") |
| "Send an email to Greta." | send_email(recipient="Greta") |
| "Go to figma.com." | open_url(url="figma.com") |
| "Send a Slack message saying 'ship it'." | slack_send(message="ship it") |
| "Create a Jira ticket titled 'broken link'." | jira_issue(summary="broken link") |
| "Compose an email to Judy and search for how tall is Everest." | send_email(recipient="Judy") + web_search(query="how tall is Everest") |
Multi-turn coding (grounds a follow-up arg from a tool response):
read_file(tests/test_api.py) → result → run_tests() → "FAILED…" → fix.
At catalog scale (100s–1000s of tools) selection is done by retrieval (top-k) instead of a
fixed head. See the LocalAgent repo.
Load (pure PyTorch, no transformers)
import json, torch
from huggingface_hub import hf_hub_download
from localagent.model import LocalAgentLM, ModelConfig
cfg_d = json.load(open(hf_hub_download("danelcsb/localagent-ultra-tiny-1m", "config.json")))
cfg = ModelConfig(**{k: v for k, v in cfg_d.items() if k in ModelConfig.__dataclass_fields__})
model = LocalAgentLM(cfg)
from safetensors.torch import load_file
model.load_state_dict(load_file(hf_hub_download("danelcsb/localagent-ultra-tiny-1m", "model.safetensors")))
model.eval()
See the LocalAgent repo for the grounded decoder / agent runtime (tool head, pointer head, retrieval, parallel-call decode).