Models & Research

    We train the models, not just the prompts.

    Most teams consume AI models. AltSlate Labs also trains them. Our Jugnu family of compact language models is built from scratch, and the same discipline shapes the custom and fine-tuned models we deliver for clients.

    Model capabilities we offer

    Pretraining from scratch

    We build models from the tokenizer up — architecture, data pipeline, and training loop — not just prompt wrappers.

    Fine-tuning & distillation

    Adapt frontier or compact models to your domain, then distill into something small enough to run cheaply in production.

    Evaluation & benchmarking

    Reproducible harnesses, honest baselines, and the eval discipline to know a model is actually better before it ships.

    The Jugnu Research Family

    An open, reproducible line of sub-150M models exploring how much capability fits into a small budget. Our 110M flagship reaches 81.25% on BLiMP, a standard test of grammatical understanding — matching a larger baseline with fewer parameters — and our 53M ranks #3 on the sub-150M Tiny-ML Leaderboard.

    These are base research models — Tiny-ML leaderboard entries under a permissive license, not instruction-tuned production assistants. Client systems use whichever model best meets their quality, cost, latency, and privacy needs.

    JugnuLM-110M

    Research
    81.25BLiMP

    matches a 125M baseline at ~12% fewer parameters

    Deep-thin Qwen3 decoder with QK-Norm, grouped-query attention, RoPE, and tied embeddings. The family flagship.

    Params
    109.7M
    Tokens
    ~8B
    Context
    2,048
    Type
    Base (R0)
    Data
    FineWeb-Edu (sample-10BT)
    View on Hugging Face

    JugnuLM-53M

    Research
    78.14BLiMP

    Our sub-150M leaderboard entry — how much capability fits into 53M parameters trained from scratch.

    Params
    53.5M
    Tokens
    ~12B
    Context
    2,048
    Type
    Base
    Data
    FineWeb-Edu (sample-10BT)
    View on Hugging Face

    JugnuLM-110M-R1

    Research

    Adds value residuals to the base recipe — testing whether richer residual pathways lift sample efficiency.

    Params
    109.7M
    Tokens
    ~8B
    Context
    2,048
    Type
    Base + value residuals
    Data
    FineWeb-Edu (sample-10BT)
    View on Hugging Face

    JugnuLM-110M-R2

    Research

    Swaps the optimizer for Muon — an ablation on how second-order-style updates change small-model training dynamics.

    Params
    109.7M
    Tokens
    ~8B
    Context
    2,048
    Type
    Base + Muon optimizer
    Data
    FineWeb-Edu (sample-10BT)
    View on Hugging Face

    JugnuLM-110M-R3

    Research

    Re-weights the training data mixture — isolating how corpus composition alone moves downstream quality.

    Params
    109.7M
    Tokens
    ~8B
    Context
    2,048
    Type
    Base + data blend
    Data
    Blended corpus
    View on Hugging Face

    Research Publications

    Peer-visible research from AltSlate Labs, published on arXiv: methods, benchmarks and negative results.

    arXiv:2610.16655 · cs.AIOct 2026

    Does a Hosted Decision Model Respect Irrelevant Alternatives? A Pre-Registered Test of Option-Set Dependence in Jev

    Dushyant Rajput, Nirdesh Chauhan, Siddharth Kosaraju

    Decision APIs return a probability per option, but those numbers describe the options only if the model ignores irrelevant alternatives. In a pre-registered study of 299,463 calls to a hosted decision model, every added option shifted the odds between the right answer and its strongest rival by 0.20–0.52 nats, growing with the log of the option count — so its probabilities are best read as relative to the list they were computed over.

    Read on arXivRead the blog post
    arXiv:2609.37832 · cs.AISep 2026

    Can a Cacheable Decision Model Follow Rules?

    Dushyant Rajput, Nirdesh Chauhan, Siddharth Kosaraju

    Encoding each candidate once and reusing it across states makes a decision model about 5× cheaper at 77 candidates, but the tested conversion lost nearly all rule-sensitivity (recall@1 1.00 → 0.24) while the joint scorer held 1.00. Targeted counterfactual supervision restores it on held-out synthetic rules; on real rules the joint scorer still wins (0.861 vs 0.500 on the short tier), so transfer is not established.

    Read on arXivRead the blog post
    arXiv:2609.01345 · cs.AISep 2026

    Cheap Verifiers, Large Blind Spots: Measuring the Reliability Cost of Cost-Saving Cascades

    Dushyant Rajput, Nirdesh Chauhan, Siddharth Kosaraju

    Inference cascades that escalate to a verifier have a blind spot that grows with student capability and shrinks only by paying the frontier price. Corrective fine-tuning on the rejected tail collapses the student, and every in-loop metric stays flat while true error climbs — the system cannot see its own degradation.

    Read on arXivRead the blog post

    Need a model tuned to your domain?

    From fine-tuning a frontier model to shipping a distilled one on your own hardware.

    All releasesDiscuss your project