Optra Research

Optra Research

AI for a billion Indians.

Small language models and fine-tuning

We train small models and fine-tune open-weight ones into AI that runs on our own GPUs, built for the work people actually do.

Write to usSee what we do

Our mission

AI for a billion Indians. Our mission is to put AI in the hands of a billion Indians. We will stop only when a billion Indians have access to AI.

Our vision

Why it hasn't happened yet

AI today is billed by the token. Every question, every task, every automation runs up a meter, so AI stays with the few who can afford it.

How we do it differently

How we're changing that

We don't stop at renting out GPUs. We build the whole stack: our own GPUs, our own fine-tuned open-weight models, and the apps people use every day: Optra Code, Optra Studio, Videora IQ, and more to come. Because we own every layer, we can price for everyone: unlimited use, no per-token bills, no surprises.

Billed per tokenUnlimited tokens

Open by design

Open by design

All our research is on open-weight models, AI that anyone can run, inspect and improve. That is how AI becomes truly local.

What we do

Small language models and fine-tuning

We fine-tune Qwen 3.8 and other models, and we train small language models, for teams that need a model built for their own work.

  • Train a small language modelA small model for one job, made for your task and your data.
  • Fine-tune a large language modelTake a large model and teach it your domain, your format and your way of answering.
  • Start with the problemTell us what the model must do and what data you have, and we will tell you what we can do.
DataBase modelLoRA / QLoRAEvaluationHosting

For enterprises

For enterprises

  • Fine-tuningAn open-weight model adapted to your domain, language and tasks.
    Qwen 3.8LFM3-1.7BFLUX.2 klein 9BQwen-Image 2.1Qwen-Image-Edit 2511Kandinsky 6.0 LiteWan 2.1Wan 2.2Wan 2.2 AnimateWan 2.2 S2VWan Animate 2Wan DancerLTX 2.5MiniMax H3Singularity H3FastH3 OmniSkyReels V3InfiniteTalkSeedVR2 7BFlashVSR 1.1ACE-Step 1.5MMAudioHunyuanVideo-FoleyThinkSoundChatterbox MultilingualKokoroVeenaIndic Parler-TTSF5-HindiIndicF5Sooktam-2Audar TurboOmniVoiceMaya 2 NativeQwen3-TTS
  • LoRA trainingSmall adapters trained on your data, quick to train and easy to swap.
  • QLoRA trainingLoRA on a quantised model, so large models train on fewer GPUs.
  • Small models for one jobA small language model trained from a larger one for your task, cheaper and faster to run.
  • Evaluation on your dataThe tuned model measured against the base model on a test set you agree, before it goes live.
  • HostingYour model served on dedicated Optra GPUs, with unlimited tokens at an affordable price.

Models on our GPUs

Models on our GPUs

These models are downloaded on the Optra GPUs. We are fine-tuning all of them.

  • Text
    Qwen 3.8LFM3-1.7B
  • Image
    FLUX.2 klein 9BQwen-Image 2.1Qwen-Image-Edit 2511Kandinsky 6.0 Lite
  • Video
    Wan 2.1Wan 2.2Wan 2.2 AnimateWan 2.2 S2VWan Animate 2Wan DancerLTX 2.5MiniMax H3Singularity H3FastH3 OmniSkyReels V3InfiniteTalk
  • Upscaling
    SeedVR2 7BFlashVSR 1.1
  • Music
    ACE-Step 1.5
  • Sound for clips
    MMAudioHunyuanVideo-FoleyThinkSound
  • Voice
    Chatterbox MultilingualKokoroVeenaIndic Parler-TTSF5-HindiIndicF5Sooktam-2Audar TurboOmniVoiceMaya 2 NativeQwen3-TTS

Our products

What we have built

Videora IQ: most security cameras record everything and detect nothing. Videora IQ watches your existing cameras in real time and alerts your team the moment something needs attention: intrusion, fire and smoke, loitering, safety-gear checks and more. No new hardware.

Built for corporate offices, hospitals, schools and campuses, factories, public safety agencies and event venues.

Contact

This is just the start.

Talk to us: [email protected]. Tell us what your team needs a model to do.

Write to us