Compiles and compresses AI models to run on the neural chips inside appliances, cars and cameras
Consumer electronics makers want speech, vision and small language models running on device, but each product line ships with a different neural chip and every vendor toolchain behaves differently. One appliance maker told us that porting a wake-word model and a vision model to three chip families took six engineers four months, and accuracy still dropped after quantisation.
Tinyfold takes a trained model, searches quantisation and pruning settings layer by layer, and compiles it for the target chip, checking accuracy against the customer's own test set at each step. Engineers see size, latency, power and accuracy side by side for 14 supported chips and export a ready-to-flash package in hours instead of months.
Device makers in Korea, Japan, Taiwan and China shipping over 1 billion products a year with neural chips; on-device AI tooling spend is projected to approach US$2B by 2029.
Annual licence for engineering teams starting at US$60K, plus a per-device royalty of a few cents once a compiled model ships in volume.
Han Seo-yeonCEO and Co-founder
Park Ji-hoonCTO and Co-founder
Choi Min-jaeCompiler LeadSeo-yeon and Ji-hoon spent years writing chip toolchains before starting Tinyfold, and it shows in how they talk about memory layouts and calibration data. Six paying device makers at seed, with one model already shipping in a robot vacuum line, is a solid start. The per-device royalty could become meaningful as volumes grow.
We used Tinyfold to move a defect-detection model from a camera module's CPU to its new 2 TOPS NPU. The layer sensitivity view showed exactly which layers to keep at higher precision, and the export package passed our firmware checks on the second try. Documentation is still thin and we needed help from their engineers twice.
Reply from Han Seo-yeon, CEO
A full guide to mixed-precision exports went live this month, written from your team's questions.
Device makers I interview complain that each chip vendor's toolchain locks them in and behaves differently. Tinyfold's side-by-side benchmarks across 14 targets give engineering and procurement teams a fair comparison for the first time. Buyers will ask whether chip vendors cooperate or treat it as a threat to their own tools.
On convolutional vision models the results are strong: four times smaller with under a point of accuracy lost. When I tried a 1B-parameter language model on the same NPUs, the accuracy drop at int4 was much larger and the search took hours. The team was candid that transformer support is early, but buyers should test their own models before committing.
Coming soon. Testimonials are invited by the founder and shown separately from open reviews.
Coming soon. Follow Tinyfold to hear about product and funding updates when this launches.
