Wednesday afternoon, I handed a client’s chat screenshot to AI
While organizing assets for a course-design client in Hangzhou, I tossed a product image, chat history, and a short description into an online model. It misread some text in the price table, and I spent ten minutes fixing it. I thought at the time: if images and text could be processed together on my own computer, would that be easier? I later realized this is not the kind of tinkering every side-hustler needs to take on.
What Inkling 2 Small actually is
Thinking Machines recently released Inkling 2 Small. It is not another model inside a typical chat window, but a multimodal model whose weights can be downloaded, run, and trained further: 594 billion total parameters, with about 17 billion activated for each calculation, and the ability to process both images and text. The company says it completed roughly 68 hours of training and outperformed comparable models on some tests.
What I care about most is customizability. Lin Xiao, an independent developer in Chengdu, is using this type of model to process product images and Chinese-language materials. He hopes eventually to retain his own workflow for several small clients instead of uploading every asset to an external platform. He also reminded me that open weights do not automatically mean easy deployment.
What replicating it costs today—in money and time
Based on the public information, training from scratch is definitely not for small teams: it requires an expensive GPU cluster, with a barrier far higher than building a website. I got this wrong before too. I initially read “only 68 hours of training” as “I can replicate it on an ordinary computer in 68 hours,” but those 68 hours depend on dedicated compute and a complicated process.
If I only want to try it, I can first check whether Tinker provides an API entry point without downloading the model. The company also plans to release the weights under Apache 2.0, but that depends on external safety testing. For someone starting from zero, the first step is to open Inkling’s page on Tinker, run one piece of text and one image, and track the cost and speed. There is no need to download the weights today. I would budget a few tens of dollars and about one to two hours for a trial. Actual local deployment may require several thousand yuan in hardware and several days of setup. The real barrier is understanding model environments and GPU configuration; knowing how to write code is not enough.
I’d decide based on where I am now
If I am just starting and mainly writing copy or handling customer replies, I will keep using mature online models. Pay-as-you-go is steadier, and chasing a new model is unnecessary. If I already have one or two clients and regularly process product images or scanned contracts, I can spend a day running a small trial and compare recognition accuracy with the cost per job. If I am scaling, already have a technical colleague, and am starting to care about data retention and custom training, then downloading the weights or contacting the official team becomes reasonable. Not trying it now is fine too. I can reassess when the weights open up, pricing becomes clearer, or an existing tool genuinely blocks my work.