Kailash Labs, a two-person Y Combinator Fall 2026 startup based in Bengaluru, says its Marlin-2B video model ranks fifth among open-source systems on TimeLens-Bench and beats OpenAI's GPT-5 by 3.9 mIoU (mean intersection over union, a measure of how well predicted time segments overlap with the correct ones) on the combined average, while being 3.5 times smaller. Those figures are the company's own and have not been independently verified.
The claim sits at the center of a larger pitch: that small, specialized models can match far bigger ones at understanding when things happen in a video.
Why small video models matter now
Running frontier video models is costly. A September 2026 arXiv survey, "Why Is Video Still So Expensive?", catalogs inference bottlenecks across the video-model pipeline.
Kailash is betting that narrow models, post-trained for a single task, can undercut those costs. According to its YC List profile, updated in October 2026, the company calls itself "a factory for small and specialized video reasoning models."
What Marlin-2B does
Kailash released Marlin-2B on May 14, 2026. Per the company's work page, it targets timestamp-grounded dense captioning (describing events in a clip along with when they occur) and natural-language temporal grounding (finding the moment a text description refers to). It is built to run fast and to be fine-tuned for particular jobs.
A post on the NemoStation blog describing the build said the model was fine-tuned from Qwen3.5-2B-Base and ships with open weights under an Apache 2.0 license. It is hosted on Hugging Face as NemoStation/Marlin-2B, and a community repository, Hook12aaa/marlin-mlx, supports local inference on Apple Silicon.
A demo site run by NemoStation uses a single L4 GPU with vLLM, an open-source inference engine. A crawl in October 2026 found roughly 15 seconds of warmup on the first request after an idle period.
Founder Shubham Sharma announced the open-source release on Twitter in May 2026, describing Marlin-2B as "a tiny VLM to extract structured information from videos."
The benchmark argument
Kailash also argues that the yardsticks themselves are flawed. In a May 26 blog post, "Part 1: The Map Was Wrong," credited to Aryan Jain, the company said its audits of the CaReBench and DREAM-1K benchmarks found ground-truth errors in 65% to 70% of items. The post contended that text-judged evaluation misleads and that video-verified ground truth matters more.
"Marlin-2B is the smallest model in the open-source top 5," Jain wrote.
TimeLens, a benchmark from ARC Lab published at CVPR 2026 in May, includes baseline numbers for GPT-4o, GPT-5 and Gemini-2.5-Flash and Pro. Its leaderboard changes over time, so Marlin-2B's current rank should be checked against the live TimeLens page.
Cost and compliance tailwinds
Hardware pricing, at least, favors small models. Compute.cafe's panel, crawled in November 2026, listed an L4 rental floor of $0.21 per GPU-hour from AIC Cloud, with provider medians between $0.32 and $0.98. A September 2026 post from Cipher Projects put A10G and L4-class instances at about $0.80 to $1.20 an hour on major clouds.
Privacy rules add another push. The EU AI Act began applying certain provisions on Aug. 2, 2026, according to the European Commission's digital strategy pages. India's Digital Personal Data Protection Rules carry staged milestones on Nov. 13, 2026, and May 13, 2027. A DHI guide updated Oct. 5, 2026, argues that sensitive video favors on-premises deployment and that timestamped, auditable outputs help with compliance logging.
Competitors and incumbents
The field is crowded. Twelve Labs raised a $100 million Series B on July 1, 2026, according to its GlobeNewswire release, and distributes models including Marengo 3.0 and Pegasus 1.5 through Amazon Bedrock. "Models commoditize. The intelligence layer that composes them does not," CEO Jae Lee said in the release.
Google offers its Video Intelligence API and Gemini 2.5 Flash and Pro with video inputs. Amazon Rekognition carries an end-of-support notice, effective April 30, 2026, for new customers on certain streaming video analysis and batch moderation features. Veritone announced Veritone Video Intelligence in September 2026, and Clarifai remains an active computer-vision API vendor.
Research rivals are also crowding the small-model niche. Apple's FastVLM, MiniCPM-V 2.6, Penguin-VL and NVIDIA's VideoITG all target efficiency or temporal grounding.
Investors, meanwhile, are paying for the category. TechCrunch reported on Aug. 31, 2026, that AI media-search startup Clipto reached a $250 million valuation.
Funding and the factory pivot
Kailash's funding is thinly documented. CB Insights lists a $0.5 million convertible note from Y Combinator dated Sept. 1, 2026, while Dealroom lists a standard $125,000 YC check. The two may reflect different instruments, and as of Nov. 2, 2026, neither had been confirmed by a company press release.
The company's homepage, crawled in November 2026, leads with a different pitch: "Robot workforce, deployed on your lines." It describes robotized vision stations for factory floors and invites visitors to book a pilot. That reads as positioning toward industrial deployments. As of the November 2026 site crawl, no customers are named on the company's homepage or YC List profile, and no headcount is available beyond the two people listed by YC List.
What comes next
McKinsey's 2026 tech trends report says physical AI will reach manufacturing and logistics first because the economics are clearer there. That is the lane Kailash's homepage targets.
The real test is whether a 2B model that scores well on a public benchmark holds up on a working production line. In a late-2026 post, the @KailashLabs account said the company was "opening up the factory for orders," according to a third-party archive of the account.
