Small models, sharp skills
Compact language models specialised for one domain, and the evaluation methods to prove they hold up — toward a blueprint for small language model design.
Deep-tech AI research for sovereign, regulated work.
23N AI Labs is an applied AI research lab in Gandhinagar, India. We study how small language models and agentic systems behave under real constraints — Indian languages, limited compute and strict data residency — and we publish what we measure.
What we build
Agents you can audit
Agentic systems that plan, use tools with scoped permissions and complete multi-step work, with full audit trails and a person signing off where it counts.
Sovereign by design
Models that run on your infrastructure or in-country data centres, so sensitive data never leaves your jurisdiction. Data residency is part of the architecture, not an afterthought.
Our lab works on applied AI's hardest constraints — and publishes the measurements.
Research focus
01 / Small language models
Small language models
- Domain-specialised models that fit modest hardware
- Distilling expertise without losing accuracy on edge cases
- Evaluation suites for regulated, high-stakes tasks
02 / Agentic AI
Reliable agents
- Reliability over long, multi-step workflows
- Verifiers and compilers as guardrails for generated work
- Human review built into the loop, not bolted on
03 / Sovereign AI
Sovereign and private deployment
- On-premise and in-country inference on CPUs and GPUs
- Data residency and auditability by default
- Honest cost and performance measurement
04 / Indian languages
AI for Indian languages
- Tokenizer efficiency and cost across scripts
- Models and evaluation for local context
- Multilingual systems for public-scale use
Recent working papers from the lab.
Latest research
Agentic AI · 16 Oct 2026
Which language keeps an LLM honest?
152 coding tasks in seven language setups: compiler feedback rescues many attempts, but every toolchain still accepts wrong programs.
Read the paper →Indian languages · 10 Oct 2026
Token counts across Indian languages
Five tokenizers, six languages and 300 aligned rows show a wide spread in token counts — a direct cost and latency gap for Indian-language AI.
Read the paper →Sovereign deployment · 10 Oct 2026
My Model, My Data, My Cloud
A measured CPU case study of running models on your own infrastructure, separating local routing performance from cost assumptions.
Read the paper →Small models · 10 Oct 2026
How far do free coding tokens go?
A 50-task local coding run measuring test passes and decode throughput for small, locally hosted models.
Read the paper →Where our research meets regulated industry.
Industries
Government
Sovereign AI for public services
Small, sovereign-deployable models and auditable agents for case work, citizen services and compliance review — running on government or in-country infrastructure so citizen data stays within national borders and residency rules.
Partner with usFinance
Private AI for regulated finance
Built by engineers who have run card issuing and payment processing at scale. Agents for operations, disputes and compliance, on models that keep financial data in-house and meet regulators' expectations.
Partner with usEducation
AI for learning at scale
Multilingual small models and assistants for schools and learning platforms, designed for Indian classrooms and for keeping student data private.
Partner with us