We build AI that works through hard problems, checks its own reasoning, and says so when it doesn't know.
Intelligence is the most human thing there is. It is how a child learns to speak, how a doctor reads a hard case, how an engineer spots the flaw before it breaks. For the first time we can build it rather than only be born with it. We want to understand it well enough to build it carefully, and then give it to everyone: so that a farmer, a nurse, a student and a founder can each think bigger than they could alone. Intelligence for everyone is the whole project of Kishiv Labs.
Bigger models will keep getting better, but size alone doesn't explain why they work or where they fail. We build our models from the ground up so that every part can be examined, measured and improved on purpose.
The next real advance in AI is not better recall. It is systems that can plan, explore many ways through a problem, test their own ideas and change course when the evidence says so. Everything we research points at this.
We don't believe the path forward is more human labels. Our methods let a model learn from its own best reasoning, so improvement scales with thinking rather than with annotation.
A system that knows what it does not know is more valuable than one that sounds sure. We hold our models to that standard: honest about their limits, willing to check or ask, and open to scrutiny.
We publish our methods and we build models that can run where their users work. Intelligence that lives behind someone else's door isn't intelligence you can rely on.
Every model here is trained by us, from the ground up, and can run where its users work. Two are available today; the third is in development.
TRIDENT is our method for improving how a model reasons. It needs no human-written answers and no extra training data.
Think of a student preparing for an exam. They try each practice question several ways, check which working holds up, spend most of their time on the questions they only sometimes get right, and then study their best solutions. TRIDENT does the same for an AI model.
Combines a guided search over many reasoning paths with a training plan that focuses on the problems a model is least consistent on. Applied to Qwen3-4B, it raises the score on graduate-level science questions from 28.28% to 42.42%.
Teaches a model by having it predict the meaning of hidden parts of an image instead of redrawing every pixel. It informs how we think about learning in terms of meaning rather than surface detail.
AI models are compared on standard tests, each with a fixed set of questions. Scores are the percentage answered correctly. Here is what each test checks.
Some tests show the model a few solved examples before it starts. "5-shot" means five examples; "0-shot" means none.
We applied TRIDENT to Qwen3-4B, a publicly available model, and tested it before and after. Scores rose on five of six tests with no human-written answers added.
| Test | Examples shown | Before | After | Change |
|---|
Accuracy %. From "TRIDENT: Thought-based Reasoning and Improvement through Deep Exploration of Neuronal Trees" (Puri, Khandelwal, Joshi, Yadav, December 2025).
M4 is our first model, trained from the ground up. It has 2 billion parameters, a rough measure of a model's size. Below it sits next to open models from 1 to 7 billion parameters. M4 scores ahead of Llama 3.2 1B on three of five tests and trails larger, more established models. Closing that gap is the job of M5.
| Model | HellaSwag | Winogrande | ARC-C | TruthfulQA | GSM8K |
|---|---|---|---|---|---|
| Kishiv M42B | 59.1 | 58.7 | 38.5 | 35.1 | 24.1 |
| Llama 3.21B | 46.5 | 58.8 | 32.5 | 44.0 | 7.5 |
| SmolLM21.7B | 62.0 | 60.4 | 42.5 | 36.0 | 32.0 |
| Gemma 22B | 71.5 | 66.3 | 48.0 | 37.0 | 23.0 |
| Llama 3.23B | 69.8 | 62.0 | 47.2 | 41.8 | 46.2 |
| Mistral7B | 81.0 | 74.0 | 53.8 | 42.2 | 37.0 |
Accuracy %. GSM8K is 5-shot.
Decider is tested on public decision benchmarks: sets of situations where the model must pick an option, answer yes or no, or give a score. Below it sits next to two purpose-built decision models of similar size. Decider leads on bev-decision and the hardest JevBench tier, and trails on the others, mostly where a question is worded in a way it hasn't seen.
| Benchmark | What it tests | KISHIV Decider 254M | LAYA 0.4B | bev-decider 0.4B |
|---|---|---|---|---|
| bev-decision | Everyday choices between options | mid-70s | ~50 | mid-70s |
| typed-decisions | Choices with a fixed answer type | high 60s | mid-70s | — |
| sysone-bench v2 | Intent, sentiment, NLI, moderation | low 60s | high 60s | ~70 |
| JevBench | Mixed judgement questions | ~50 | high 50s | mid-60s |
| JevBench, hard tier | The hardest JevBench questions | high 30s | mid-30s | high 40s |
LAYA and bev-decider figures are from their published model cards. Decider was scored on the same public items with each harness's own rules. Highlighted cells mark the best in each row.
Our advisers are experienced technology and business leaders. They help us choose which research problems are worth solving, test our thinking before we commit to it, and connect our work with the people and organizations who can use it.
Engineering leader at AppDynamics, Cisco's application performance and observability business, where he runs large-scale product engineering. Advises Kishiv on building and scaling engineering organisations.
LinkedIn ↗
Co-founder and CTO of Workorb AI. Previously co-founded Sysomos (acquired by Marketwired) and Aislelabs (acquired by Constellation Software). PhD in computer science, University of Toronto; B.Tech, IIT Bombay.
LinkedIn ↗
Former CTO of Manthan, the retail analytics and AI company, after senior engineering roles at Yahoo, Sun Microsystems, Motorola and Good Technology. Chaired the NASSCOM Product Council.
LinkedIn ↗
Former Vice President at Samsung. Co-founder and chief scientist of Tekk.ai and CTO of INDIC EMS Electronics. Earlier built the control and data software for the CARMA radio telescope array.
LinkedIn ↗
connect@kishiv.in