hcawn-fc

A tiny function-calling model (5.7M parameters, byte-level, CPU only) for 7 phone actions: send_email, create_calendar_event, create_contact, show_map, open_wifi_settings, turn_on_flashlight, turn_off_flashlight.

It runs locally in well under a second, always returns valid JSON, and holds back calls it is not sure about. Try it in the browser first: https://cawnai.com

Scope: a narrow model for a known tool set, not a general function-calling model. It does not take arbitrary tool schemas.

Install

pip install hcawn-fc
hcawn-fc doctor
hcawn-fc "Add a calendar event Team Sync tomorrow at 2pm"

Python 3.9+, any OS. Until you download the weights, hcawn-fc calls the hosted API (free demo). The local CPU engine ships for Linux x86-64, see "Run it locally".

Run it locally

The local engine runs on your CPU with no network. The weights (6 MB int8, 3.9 MB int4, CC BY 4.0) are hosted on Hugging Face and downloaded once:

hcawn-fc download            # int8 (default); also: hcawn-fc download int4 | all
hcawn-fc doctor              # shows the engine, the weights and a self-test

They are saved to ~/.cache/hcawn_fc/ and checked against a SHA-256 sum. You can also pass your own file with --model-path / Local(model_path=...).

Use it from Python

from hcawn_fc import Local

m = Local()                                   # needs the engine (Linux x86-64) and `hcawn-fc download`
r = m.call("Email dana@example.com subject Lunch body Are we still on for noon?")

r.status        # "ok" | "held" | "none"
r.calls         # [Call(name="send_email", arguments={...})]
r.confidence    # 0.0 - 1.0
r.latency_ms

Statuses:

status	meaning
ok	the call is confident enough to act on (r.calls)
held	the model proposed a call but is not sure; ask the user first (r.held_calls)
none	none of the 7 actions fits the request

Relative dates ("tomorrow", "next Monday") are computed from now (default: local time). Pass m.call(q, now=datetime(...)) to override. Local(model="int4") uses the smaller file; Local(min_conf=0.95) makes the gate stricter. One Local instance handles one request at a time (calls are serialized); use one instance per process for parallelism.

Hosted API

from hcawn_fc import Client
Client().call("Turn on the flashlight")        # https://cawnai.com, free demo, rate limited. Works on any OS.

or hcawn-fc --api "Turn on the flashlight". Raises RateLimited / ServerBusy / ApiError. Need higher limits or an SLA? Contact us (below).

Command line

hcawn-fc "Show me Paris on the map"      one request
hcawn-fc chat                            interactive
hcawn-fc doctor                          what engine / weights / CPU variant is in use
  --api  --model int4  --json  --min-conf 0.95  --model-path file.bin

The engine ships in three builds (AVX2, SSE4.1, generic) and the fastest one your CPU supports is picked at start-up. Override with HCAWN_FC_ISA=avx2|sse41|generic.

How good is it

Strict exact match (tool names and every argument value) on the Mobile Actions eval set, 961 rows, about ±2 points of noise. These numbers are for model plus the deterministic layers around it (grammar-constrained decoding, grounding of values in the query, date/time rules, keyword triggers, confidence gate):

	Exact
int8	90.4%
int4	90.3%
int8 on rephrased queries (905 rows, paraphrased by a different LLM)	82.8%
model + grammar only, no other layers (our own evaluator)	86.5%

The first three rows were measured with this package, and you can repeat them (see "Reproduce the numbers" below). Queries are limited to 300 characters; 46 of the 961 benchmark requests are longer, and on the 915 up to 300 characters int8 gets 90.3%.

Weaker on unusual phrasings ("tell him that ...", "... saying X"). The confidence gate is not a guarantee: it stops most, not all, wrong calls.

Model details

Parameters	5,742,720 (5.74M)
Architecture	hCAWN hybrid: 6 layers (4 recurrent associative-memory layers and 2 local-attention layers), hidden size 128, 4 heads, local attention window of 4,096 bytes
Vocabulary	258 symbols: 256 raw UTF-8 bytes plus start and end markers. No tokenizer
Files	int8 6.0 MB, int4 3.9 MB
Pretraining	next-byte prediction on English web text (FinePDFs-Edu, DCLM, FineWeb-Edu), no function-calling data. Best checkpoint at step 241,000 of 300,000 (about 15.8 billion bytes), validation loss 1.068 nats per byte on WikiText-103
Fine-tuning	supervised, on Mobile Actions plus 16,000 LLM paraphrases of its requests. Examples of up to 2,048 bytes (tool definitions, date, request and answer)
Speed	median about 0.2 s (int8) and 0.3 s (int4) per request on one core of a desktop CPU with AVX2

Why 300 characters? It is a safety limit of the package and the demo, not a limit of the model. The tool definitions and the date take about 1.2 KB of the 2,048 bytes the model was fine-tuned on, so there are roughly 850 bytes left for the request and the answer. In the benchmark 4.8% of the requests are longer than 300 characters (up to 740), and the model gets 43 of those 46 right.

Full details: https://huggingface.co/datatab/hcawn-fc7

Reproduce the numbers

The folder eval_check/ has the evaluation requests and a short script that uses only this package and its own scorer (strict exact match):

pip install hcawn-fc
hcawn-fc download int8 && hcawn-fc download int4
cd eval_check
python3 verify_fc7.py data/fc_mobile_eval.jsonl --model int8 --procs 4 --allow-long
python3 verify_fc7.py data/fc_mobile_eval.jsonl --model int4 --procs 4 --allow-long
python3 verify_fc7.py data/fc_mobile_eval_para.jsonl --model int8 --procs 4 --allow-long --name rewritten

Expected: 90.4% (int8), 90.3% (int4) and 82.8% (rewritten requests), within about 2 points of noise. Every miss is written to out/*_errors.jsonl. The Linux x86-64 engine is needed.

Data and licences

• Code: Apache-2.0. The native inference engine ships as a compiled binary inside the Linux x86-64 wheels; its source code is not published.
• Trained on the Mobile Actions dataset (Google, CC BY 4.0) plus query-form augmentation and 16,000 paraphrases of the training queries written by an LLM (Gemini 2.5 Flash Lite) and filtered automatically. Evaluation paraphrases were written by GPT-4o-mini and were never used for training.
• Model weights: CC BY 4.0 (Hugging Face: https://huggingface.co/datatab/hcawn-fc7). See NOTICE.
• Evaluation data in eval_check/data/: requests and expected calls derived from the Mobile Actions dataset (Google, CC BY 4.0), and 905 rewrites of them written by GPT-4o-mini.

Custom models

We build small CPU-only models like this one for your tools and your phrasing, and run them on-device or on your own servers. Contact: https://cawnai.com (contact form at the bottom).

Serbian (Latin script)

A separate model understands Serbian requests (Latin script, with or without diacritics, typos included). It is not the English model: it was fine-tuned on Serbian only and is not meant for English.

hcawn-fc download sr-int8        # or sr-int4 (3.9 MB)
hcawn-fc --model sr-int8 "upali lampu"
hcawn-fc --model sr-int8 "zakazi sastanak sa Markom sutra u 15h"

from hcawn_fc import Local

m = Local(model="sr-int8")
r = m.call("dodaj kontakt Marko Petrovic 064 123 4567")
print(r.status, r.calls)

On 450 requests rewritten as people type them (no diacritics, typos, lowercase), sr-int8 gets 85.1% exact match and sr-int4 84.9%. At the default confidence gate (0.9) the model answers about 80% of requests and about 97% of those answers are exactly right. The weakest tool is show_map. The weights are in the Hugging Face repo datatab/hcawn-fc7-sr.