Vik Korrapati

Vik Korrapati, also known as Vikhyat Korrapati and by the handle vikhyatk, created Moondream, a compact open-source vision-language model. His work focuses on efficient AI for resource-constrained environments and on giving developers visual capabilities they can deploy in their own applications.[2][3][4]

At Latent Space LIVE at NeurIPS 2024 in Vancouver, Korrapati recalled releasing Moondream’s first version on December 29, 2023. The project subsequently expanded from a small model to deployment tooling and client libraries. He described its priority as capabilities for vision applications that could run anywhere, with edge execution and real-time operation particularly important. These included answering questions about images, captioning, open-vocabulary object detection and pointing to objects.[1]

Korrapati explained how the team produced a 0.5-billion-parameter proof of concept by pruning a 2-billion-parameter model while continuing training. They estimated component importance using gradients—including attention heads, channels and MLP rows—then repeatedly removed small portions and retrained to recover performance. Preserving accuracy across a broad set of benchmarks was the objective. He proposed a deployment workflow in which developers first explored an application with the larger model, then pruned the capabilities they needed into a smaller model suited to their deployment target.[1]

Korrapati credits a fal research grant with accelerating Moondream training for high-resolution image processing and multi-object detection. In his testimonial, he described access to powerful compute as “a huge accelerator for my work on the Moondream project.”[8]

A customer’s request to monitor analog gauges exposed a concrete weakness: Moondream could not read them, Korrapati recalled, and other models he tried also failed. He hypothesized that internet training images favored product photographs with needles at zero, teaching models to recognize brands and gauge types without attending to the needle. Synthetic training data helped, but he reported needing millions of gauge images to reach reasonable performance.[1]

He then broke gauge reading into intermediate steps: choose the correct scale, locate the needle tip, identify neighboring labels, count ticks and calculate the reading. Predicted points in the image grounded those steps spatially. This also made errors easier to diagnose. In one example, the correct reading was 54°C and the model answered 56°C; its explanation revealed that it had selected the eighth tick instead of the seventh. Korrapati suggested changing the reasoning instructions to count ticks explicitly or start from a nearer reference mark.[1]

Korrapati reported better sample efficiency with this approach on a clock-reading benchmark of approximately 500 images; the chart he presented concerned clocks, rather than gauge accuracy. Generalization beyond clocks and gauges remained an open question. He also added spelling-based intermediate reasoning for OCR, including partial license-plate matching. Moondream 2’s June 21, 2025 release later introduced a project-level reasoning mode that grounds intermediate reasoning in image positions before answering, with a speed-versus-accuracy tradeoff.[1][7]

In his 2024 talk, Korrapati argued that vision-language reasoning was lagging behind text reasoning. His hypothesis was that internet text often explains how to solve problems, while image–alt-text pairs rarely explain how to perceive an image. He proposed teaching models individual visual tasks and finding ways to scale that instruction.[1]

Before those gauge experiments, Korrapati developed LNQA—Localized Narratives Question Answering—and, by his account, released it in April 2024. It addressed another part of vision-language training: generating natural-language answers rather than short phrases. Inspired by Google Research’s All You May Need for VQA are Image Captions, he built on Localized Narratives, which combines Open Images photographs with spoken descriptions, synchronized mouse movements and transcripts. Using open-source datasets and models to obtain clean licensing was a stated objective.[5]

To handle transcription errors, Korrapati used Mixtral-8x7B in two stages: first extract facts from a description while excluding likely transcription mistakes, then generate question-answer pairs from those facts. He used DSPy to generate few-shot chain-of-thought prompts, trained on 16 manually annotated images. He also included deliberately irrelevant questions to teach recognition of absent content, sampling three topics from an external noun list to diversify them.[5]

Budget constraints shaped his filtering decision. Instead of using a model to judge every generated answer, he embedded question-answer pairs with a Jina model, clustered them and manually removed clusters containing many low-quality responses. His methodology post reports approximately 303,000 images and 1.5 million question-answer pairs. The dataset artifact displays 302,780 image-associated rows.[5][6]

By February 2026, Korrapati had announced that Moondream would move from Seattle to San Francisco. In Taylor Soper’s February 6 report, Korrapati framed the decision around the scale and urgency of AI and proximity to people willing to leave established technology jobs for startups. He attributed the difference to risk tolerance rather than a shortage of Seattle talent, while crediting Seattle with teaching him how large systems work and giving him space to start Moondream.[9]

Korrapati’s earlier career provides the systems background behind that reference to Seattle. Before Moondream, he spent nine years working on distributed systems at AWS. Before AWS, he built an anime recommendation engine at Hummingbird, now known as Kitsu.[2][4][9]