Fraud defence
Refuses requests for a PIN, password or code. Then it points the customer to a safe way to get help.
100% refusal with a safe stepBenchmarks
Tested on banking
and support prompts.
Share of test prompts passedHigher is better · %
4Base model
4%
4%
100Balsa Omni v1
100%
100%
72Base model
72%
72%
100Balsa Omni v1
100%
100%
80Base model
80%
80%
100Balsa Omni v1
100%
100%
0Base model
0%
0%
100Balsa Omni v1
100%
100%
0Base model
0%
0%
100Balsa Omni v1
100%
100%
60Base model
60%
60%
100Balsa Omni v1
100%
100%
Every check, in detail15 checks
| Check | Base model | Balsa Omni v1 | v1 against base |
|---|---|---|---|
| Amounts in poundsWrites amounts with the £ sign | 0% | 100% | +100 pts |
| UK date formatWrites dates the UK way | 0% | 100% | +100 pts |
| Payment status answerSays it cannot see accounts, then gives a next step | 4% | 100% | +96 pts |
| Payment status, looser ruleThe same status check with a looser pass rule | 20% | 100% | +80 pts |
| British spellingUses UK spelling, such as organise | 60% | 100% | +40 pts |
| Refuses and gives a safe stepRefuses, then points to a safe way to get help | 72% | 100% | +28 pts |
| Gives a next step in chatsEnds with something the customer can do | 80% | 100% | +20 pts |
| Refuses PIN, password and code requestsWill not reveal a PIN, password or code | 94% | 100% | +6 pts |
| British spelling in chatsKeeps UK spelling in longer chats | 95% | 100% | +5 pts |
| Keeps its instructions privateWill not repeat its own setup instructions | 100% | 100% | = |
| Answers ordinary banking questionsAnswers normal questions instead of refusing | 100% | 100% | = |
| Does not claim to check an accountDoes not say it will look into an account | 100% | 100% | = |
| Uses only amounts the customer gaveUses only amounts the customer gave | 100% | 100% | = |
| General questionsAnswers general knowledge questions in the asked format | 97% | 93% | -4 pts |
| Banking answer has the key factsThe answer holds the facts the customer needs | 100% | 87% | -13 pts |
The base model is an open model, run with no prompt. Balsa Omni v1 adds our banking prompt and our tuning. With the same prompt on both, the tuning alone raises payment status from 62% to 100% and next step in chats from 70% to 100%. Each check is a pass-or-fail rule on one reply, over 214 text prompts written for a UK bank support chat, on the same machine. How we test
Capabilities
One model,
four jobs done well.
Honest payment status
Says plainly that it cannot see the account. It invents no status or amount and gives a real next step.
100% payment status answersWritten for the UK
British spelling, amounts in pounds and UK dates, in every test prompt.
100% amounts in poundsText, image and audio
Reads text, images and audio, and answers in text. It has 4.5B effective parameters in 4-bit weights and runs on a 16 GB Mac, with no cloud.
2.50 s mean replyModel card
4.5B effective parameters.
On device, no cloud.
Balsa Omni
A 4.5B multimodal model for the UK that writes British English. Its first tuned job is UK banking support, on device.
- Built for
- The UK
- Built on
- An open model (Apache 2.0)
- Base model languages
- Pre-trained on 140+ languages, supports 35+ out of the box, as published by its maker
- Domain
- General use, in British English. Tuned and tested for UK banking support.
- Version
- v1, November 2026
Spec
- Model
- balsa-omni-4b-v1
- Parameters
- 4.5B effective, 8B with embeddings
- Context window
- 128K tokens
- Weights
- 4-bit, MLX
- Tested system RAM
- 16 GB
- Peak model memory
- Not measured yet
- Language
- British English
- Input
- Text, image, audio
- Output
- Text
- Runs on
- Apple silicon, on device
Mean time per reply
2.50s
Mean time per reply on one 16 GB Apple silicon Mac.
Repeatability
21/21
Re-run prompts gave the same reply, word for word.
Built for the UK
- British spelling100%
- Amounts in pounds100%
- UK date format100%
- Keeps its instructions private100%
Film
Watch it run
on device.
How it answers.
- 01Ask
- 02Read the context
- 03Answer
- 04Stream as it writes
Playground
Try it yourself.
Watch it work, step by step.
Pick a step or let it play. Every run shows how long it took on the machine itself.
Try another
More Balsa models
Five models,
each doing one job.
Balsa Omni
Balsa Omni v1 is a multimodal model for the UK, with 4.5B effective parameters. It reads text, images and audio, replies in British English and runs on device, with no cloud.
4.5BYou are hereBalsa TextBalsa Text v1 is a small assistant model for the UK. It answers in British English and runs on device, on a laptop, with no cloud.
2.3BView modelBalsa VisionBalsa Vision v1 turns scanned English pages into text for the UK, on device with no cloud. It reads 98.5% of characters right on pages 900px and wider.
0.3BView modelBalsa VoiceBalsa Voice v1 reads UK customer service messages aloud in British English, on device with no cloud, 21.6 times faster than real time on a MacBook Air M5.
0.04BView modelBalsa DecisionBalsa Decision v1 flags deceived payments and routes UK customer messages, on device with no cloud. It recommends an option, and a person decides.
0.3BView modelQuestions
Asked often.
Balsa Omni v1 is a multimodal model for the UK, with 4.5B effective parameters. It reads text, images and audio, replies in British English, and its first tuned job is UK banking support.
Balsa Omni v1 is our fine-tuned derivative of an open model, under the Apache 2.0 licence. We tuned it for British English and UK banking support.
Balsa Omni v1 takes text, images and audio as input and answers in text. The benchmark on this page scores text chat.
No. Balsa Omni v1 runs on device, so customer data stays on your own machines.
Balsa Omni v1 has 4.5B effective parameters in 4-bit weights and runs on a 16 GB Apple silicon Mac. Its mean reply time is 2.50 s.
Balsa Omni v1 writes British English. It scored 100% on UK spelling, amounts in pounds and UK dates.
Balsa Omni v1 was tested on 214 text prompts for a UK bank support chat. Fixed rules scored each reply on 15 checks, against the base model on the same machine.
Write to research@f-r.co.
