Xantyr · Own your model

Every prompt is a meter you don't own.

Spend · this month

$31,000

▲ climbing in real time

Claims · this month

1.5M

each one sent to a third party

Projected · annual

$564K

for one task, rented forever

Spend trend · 30 days

$47K/mo

30 days ago · ~$8K

Spend by task · one job dominates

claims intake78%

policy Q&A14%

claim summary8%

78% of the bill is one narrow task,
the exact case a small owned model wins.

XΛNTYR

TRAIN · OWN · DEPLOY

01Point it at your data.

Labeled examples

12,400

a sample of your own history

Sources

3

the data those calls were leaking

Uploaded to cloud

0%

stays on your machine

Sources · training-ready

Source	Records	Status
claims_2024/	8,200	✓ structured
policies/	3,100	✓ structured
intake_notes/	1,100	✓ labeled
coverage check	3 gaps	⚑ flagged

Pre-processing · on your machine

▸ de-duplicated −1,840 dupes

▸ PII masked & redacted 100%

▸ tokenized · split 90 / 10 ✓ ready

Dataset composition

12.4K

examples

claims · 66% · 8,200

policies · 25% · 3,100

intake · 9% · 1,100

balanced across your real case mix · ~214 tokens / example · illustrative

02Trains in minutes. No cloud.

Epoch

3/3

Loss

0.21

Elapsed

6min

GPU · local

78%

Run log

› base · llama-3.2-3b · LoRA r=16

› batch 16 · lr 2e-4 · seq 2048

› epoch 1/3 · loss 0.58 ✓

› epoch 2/3 · loss 0.34 ✓

› epoch 3/3 · loss 0.21 ✓

› eval · 92% pass on held-out ✓

› packaged → acme-intake-v1.gguf

throughput 1,180 tok/s

peak VRAM 14.2 GB

params 3.8B

Training loss · 1.00 → 0.21

1.00

0.60

0.21

epochs 1 · 2 · 3

Pass rate · held-out set

0%

Only ships when it beats your bar.

+51 pts vs the base model · 0 regressions on your cases · graded on 240 held-out examples

240 held-out cases · your model vs a general model · illustrative

your model · passes your bar

a general model · misses your cases

Task	Your model	General
claims intake	96%	44%
policy Q&A	93%	52%
claim summary	89%	38%

03A file you own.

Model size

4.2 GB

portable · one file

Format

.gguf

runs anywhere

License

Yours

no per-seat, ever

Egress

0 bytes

stays in your env

The artifact

◈

acme-intake-v1.gguf4.2 GB · trained on your 12,400 examples

Package includes

▸ fine-tuned weights merged

▸ tokenizer + config bundled

▸ evaluation report 92% pass

▸ lineage manifest signed

weights exportableno lock-inre-train anytime

Deploy anywhere · same file

Local machine

a laptop or a single workstation

✓

On-prem server

your datacenter, your hardware

✓

Air-gapped network

no internet connection at all

✓

Private cloud (your VPC)

your account, your keys

✓

Managed by your MSP

your partner runs it for you

✓

one artifact. Your environment, your rules.

In production. Nothing leaves.

Claims handled · today

0

≈ 1.5M this month · the same calls

p50 latency

0.9s

on your own hardware

Cost · per 1k calls

$0.01

frontier API was ~$30

Data egress

0 bytes

nothing left the building

Throughput · last 60 min

req/min

peak 1,040

Request stream

14:26:02 claims-intake · 200 · 41 tok/s · local

14:26:03 claims-intake · 200 · 39 tok/s · local

14:26:03 policy-qa · 200 · 44 tok/s · local

14:26:04 claim-summary · 200 · 42 tok/s · local

14:26:05 claims-intake · 200 · 38 tok/s · local

14:26:05 policy-qa · 200 · 45 tok/s · local

14:26:06 claim-summary · 200 · 40 tok/s · local

14:26:06 egress check · 0 bytes sent · ✓

streaming · 6 worker nodes · 100% local · 0 bytes egress this session

network: connected

OFFLINE

Your data never leaves.

you › ask it again, internet off this time

watch the connection, then watch it keep working

Auditable by design.

Model registry

Version	Trained on	Size	Weights
acme-intake-v1 · current	claims + policies	4.2 GB	export ✓
acme-intake-v0.9	claims_2024	4.2 GB	export ✓
acme-intake-v0.8	claims (partial)	4.1 GB	export ✓
acme-intake-v0.7	pilot subset	4.1 GB	export ✓

Lineage · v1

acme-intake-v1 ← claims_2024 + policies
← 12,400 records · sha256 a1f7…9c2e
every version is a file you can re-download.

full lineageweights exportableno lock-in

Audit log

ingested · 12,400 records · on-device · 14:00

trained · v1 · 3 epochs · local GPU · 14:02

evaluated · 92% pass · held-out · 14:19

deployed · on-prem node-1 · 14:26

signed · lineage hash recorded · 14:26

exported · weights → your storage · 14:27

Compliance posture

data residency · in your environment ✓

weights · exportable, no lock-in ✓

audit trail · full lineage recorded ✓

access · your infrastructure only ✓

retention · you set it, nothing phones home ✓

Same task, same volume. One of them you keep.

$0K

back in your pocket, every single year.

Rented $47K/mo

Owned ~$500/mo

Pays for itself in under a month·then flat, forever·figures illustrative

XΛNTYR

Own your model → xantyr.com

Request early access ▶

Own your model.