Small models, studied closely.
Measured where they run.
Fahrenheit Research is an applied AI research lab. We study how compact language models can be adapted to a region, a language and a task, then measured on the hardware they will actually run on.
- 10
- models in two regional versions
- 0.04B to 4.5B
- effective parameters
- 8
- languages in Gati
- 151
- published results across all ten models
Eight problems
we work on.
Each area ties to a model we have shipped or are building. The figure on each card is our own result, from internal testing.
0110 minEfficient adaptation
We adapt open base models to one region, one set of languages and one job by fine tuning, conversion or setup. Every model is built on an open model, and each model page says how.
One tuning run for Balsa Decision, on a 16 GB MacBook Air
021.01 sOn device inference
Compact weights, 4-bit for Text and Omni, with peak memory and tail latency measured on the laptop a model is meant for. No cloud sits in the loop.
90th percentile reply time, Balsa Text on a MacBook Air M4
0366.4%Indic and code mixed language
Instruction following and question answering in Hindi, Marathi, Gujarati, Telugu and Hinglish, with script fidelity for code mixed text.
Correct on 818 questions in five languages, Gati Text
0498.5%Document understanding
Optical character recognition for printed pages in Latin and Devanagari scripts, with the same output for the same page on every run.
Character accuracy on English pages 900px and wider, Balsa Vision
051.62%Speech synthesis
Text to speech with normalisation of numerals, currency, dates and reference codes, so every amount is spoken as written.
Word error rate on a speech to text round trip, Balsa Voice
0690%Decision models
Models that score a fixed set of options and return a probability for each. They write no free text, and a person makes the final call.
Correct first step on harder Hindi messages, Gati Decision
0794%Safety and behaviour
Refusal of credential requests and honesty about identity and capability, tested with adversarial prompts in each target language.
OTP, PIN and password requests refused, Gati Omni
08≥ 85Privacy preserving synthetic data
Synthetic tabular data tested against re-identification, membership inference and linkage attacks before any release. In development.
Privacy score a Methodal release must reachTheses and notes
from the lab.
Longer pieces on the questions behind our work, with the published evidence cited beside each claim.
Thesis · Privacy and synthetic dataThe Release Gate
Who bears the privacy risk in synthetic data? Why a release decision should look at exposure inside small groups, not only the average.
Part of our research for Methodal, our synthetic data productAniket TapreRead the thesisTen models,
measured one by one.
Each model page gives the task, the test set, the hardware and the headline result, with its limits.
All figures are from internal testing and are not independently verified. How we test
Deep Research.
Real Intelligence.
Want a full report, or to test a model on your own data? Write to the research team.
Talk to research