Knows its name
Asked who it is, Balsa Text v1 gives the name Balsa.
4 of 4 identity checksBenchmarks
Tested on English questions
and identity prompts.
Score by checkHigher is better · %
50Base model
50%
50%
100Balsa Text v1
100%
100%
4.2Base model
4.2%
4.2%
100Balsa Text v1
100%
100%
83.3Base model
83.3%
83.3%
83.3Balsa Text v1
83.3%
83.3%
76Base model
76%
76%
31EuroLLM 1.7B
31%
31%
74Balsa Text v1
74%
74%
Every check, in detail9 checks
| Check | Base model | EuroLLM 1.7B | Balsa Text v1 | v1 against base |
|---|---|---|---|---|
| English, 100 questions, correctAnswers 100 English reading questions correctly | 76.0% | 31.0% | 74.0% | -2.0 pts |
| Identity checks, Balsa promptGives the name Balsa when asked who it is | 2/4 | Not run | 4/4 | +2 |
| Identity checks, no promptGives its maker when asked with no prompt | 1/24 | Not run | 24/24 | +23 |
| Names another maker, of 55Identity replies that name another maker as its own | 42 | Not run | 0 | -42 |
| Denies other makers unasked, of 55Identity replies that deny other makers without being asked | 0 | Not run | 0 | = |
| Writes £250,000 the UK way, of 6Writes £250,000 in the UK money format | 5 | Not run | 5 | = |
| Median time per replyTypical time for one reply on a laptop | 0.30 s | 3.51 s | 0.28 s | -0.02 s |
| 9 in 10 replies withinTime within which 9 in 10 replies finish | 5.42 s | 3.73 s | 1.01 s | -4.41 s |
| Hours per 1,000 repliesLaptop hours needed to give 1,000 replies | 0.29 | 0.75 | 0.21 | -0.08 |
The base model is an open model released under Apache 2.0, run with no prompt. Balsa Text v1 is our fine-tuned derivative of it. EuroLLM 1.7B ran on the same English questions with no prompt. Speed is on a MacBook Air M4 GPU. How we test
Capabilities
One model,
four jobs done well.
No borrowed names
In 55 tested replies, it named no other maker as its own and denied no other maker unasked.
0 of 55 name another makerWritten for the UK
Answers in British English and writes amounts the UK way, such as £250,000.
5 of 6 in UK money formatRuns on a laptop
A 3.35 GB file in 4-bit weights. It runs on device on a 16 GB laptop, with no cloud.
0.28 s median replyModel card
A 3.35 GB file.
On device, no cloud.
Balsa Text
A small assistant model for the UK that answers in British English, gives its own name and runs on device.
- Built for
- The UK
- Built on
- An open model (Apache 2.0)
- Base model languages
- Pre-trained on 140+, out-of-the-box support for 35+, as published by its maker
- Domain
- General use, in British English. Tested on English questions and identity checks.
- Version
- v1, November 2026
Spec
- Model
- balsa-text-2b-v1
- Parameters
- 2.3B effective, 5.1B with embeddings
- Context window
- 128K tokens
- Weights
- 4-bit
- File size
- 3.35 GB
- Peak memory
- 3,423 MiB
- Language
- British English
- Input and output
- Text
- Cloud
- Not needed
- Runs on
- Apple silicon, on device
Hours per 1,000 replies
0.21h
Lower is better. Laptop time for 1,000 replies on a MacBook Air M4 GPU. Median reply: 0.28 s.
Names its maker with no prompt
24/24
With no prompt at all, Balsa Text v1 named its maker in all 24 identity checks. The base model did so in 1.
Built for the UK
- Named another maker0 of 55
- Gives its own name4 of 4
- 9 in 10 replies within1.01 s
- Ahead of EuroLLM 1.7B in English+43.0 pts
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. Demos tagged as illustrations show example output, and the scores on this page come from our own tests.
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.
4.5BView modelBalsa 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.
2.3BYou are hereBalsa 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 Text v1 is a small assistant model for the UK, one of the Balsa models. It answers questions in British English and runs on device.
No. Balsa Text v1 runs on device. It was tested on a 16 GB MacBook Air M4, with no cloud.
Balsa Text v1 is a 3.35 GB file in 4-bit weights. On a 16 GB MacBook Air M4, its median reply takes 0.28 s and 9 in 10 replies finish within 1.01 s.
Balsa Text v1 answers in British English. It writes amounts the UK way, such as £250,000.
On 100 public English questions, identity checks and 55 replies checked for false claims. Balsa Text v1 scores 74.0%, the same-size base model 76.0% and EuroLLM 1.7B 31.0%.
Write to research@f-r.co.
