--- title: "How Do You Access Inkling 2 Small and What Does It Cost?" url: "https://goremy.ai/blog/articles/inkling-2-small-api-pricing-access-cost" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-11-10T03:00:17.150Z" updated: "2026-11-10T17:03:03.371Z" topic: AI Pricing & Costs description: "Inkling 2 Small is available only through the Thinking Machines console with paid API credits. Here is what access looks like and what the $426 run shows." --- # How Do You Access Inkling 2 Small and What Does It Cost? > Inkling 2 Small is available only through the Thinking Machines console with paid API credits. Here is what access looks like and what the $426 run shows. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-11-10 ## How do you access Inkling 2 Small and what does it cost? You access Inkling 2 Small through the Thinking Machines console. You create an account, generate an API key, and add paid credits. There is no downloadable checkpoint at launch. A hands-on reviewer described the model as expensive. The clearest cost reference is a demo where the model autonomously fine-tuned a smaller model: about four hours of work and $426 in spend. A per-token price list was not part of the material this article draws on. ## TL;DR - Inkling 2 Small is **console and API only**, with no open weights available at the time of the review. - You need an **API key and prepaid credits** before you can send a single request. - The reviewer called the model **not cheap**, but no per-token rates were shown, so check the console for current pricing. - The one concrete cost data point is an autonomous fine-tuning job that ran **about four hours and cost $426**. - The model is a **594B-parameter mixture of experts with 17B active**, which helps explain why it is not a bargain-tier option. - It handles **text and images but not audio**, so factor that in before you commit budget. ## What is Inkling 2 Small? Inkling 2 Small is the second-generation frontier model from Thinking Machines. It is a mixture of experts model with 594 billion total parameters and 17 billion active per token. It reads text and images. The lab puts the most emphasis on coding and long-running agent work, though it positions the model as general purpose. The "active parameters" figure matters for cost thinking. A mixture of experts model only runs a slice of its weights for each token, so per-token compute is closer to a 17B dense model than a 594B one. The full weights still have to be hosted, which is a serving burden that the provider has to recover somewhere. ## How do you get access to Inkling 2 Small? Access follows the usual hosted-model pattern, with one notable restriction: the console is the only way in. 1. Sign in to the Thinking Machines console. 2. Generate an API key. 3. Add credits to your account. 4. Call the model from your own code, or from a tool that accepts a custom API key. The reviewer who tested the model followed exactly this path. They sent text prompts and also passed in an image (a hand-drawn circuit) through the API. So image input works over the same API as text. There is no mention of availability through third-party model hosts or cloud marketplaces in the source material. If you need one of those routes for compliance or billing reasons, verify it directly rather than assuming it exists. ## Is Inkling 2 Small open weights? Not at the time of the review. The reviewer said the console-only availability "is not really good enough" and speculated that the lab might eventually release something under a permissive license such as Apache 2. That was a hope, not an announcement. Treat the model as a hosted service for now. This has practical consequences: - You cannot run it locally or on your own GPUs. - You cannot quantize it or self-host it to cut per-request costs. - Your usage depends on the provider's uptime, rate limits, and pricing changes. - Data handling falls under the provider's terms rather than your own infrastructure. ## How expensive is Inkling 2 Small? The reviewer's verdict was blunt: the model "is not cheap." They put credits into the console to run their tests and flagged cost as the main warning. The source does not list input or output token prices, so this article will not guess at them. The console is the authority on current rates. What you can reason about is how spend scales. Agentic and coding work burns far more tokens than a chat exchange. A model that plans, writes code, runs it, reads errors, and tries again will make many calls with growing context. If the per-token price is high, long-running agent tasks multiply it quickly. ## What did the $426 fine-tuning run involve? This is the best cost anchor available. In a demo from the lab, the model received one job: train a small model to forecast four things (electricity demand, wind power, air quality, and Wikipedia page views). From there it worked without step-by-step human guidance. According to the description, it: - Wrote its own evaluator. - Built the training data. - Ran fine-tuning. - Tried reinforcement learning. - Debugged problems along the way. After about four hours and $426, the forecasting score had more than doubled. The arithmetic works out to roughly $106 per hour of autonomous work. The source does not break down how much of the $426 went to Inkling's own tokens versus compute for training the smaller model, so you cannot treat $106 an hour as the model's API burn rate. It is better read as the all-in price of one end-to-end project of this kind. ## Is the $426 run worth it? That depends on what you compare it to. A fine-tuning pipeline normally needs someone to design an eval, assemble data, pick hyperparameters, and babysit runs. If an engineer would spend days on that, a four-hour run at $426 is an easy trade. If you just want a decent answer to a question, it is not. Some caveats apply: - It is a single demo from the lab, not an independent benchmark. - "Score more than doubled" depends on the evaluator the model wrote for itself. - Results on your own data and tasks will differ. - A failed or wandering run still costs money. If you try something similar, set a hard spending cap on your credits and watch the first run closely. ## How does Inkling 2 Small perform for the price? The benchmark picture the reviewer described is mixed. A DeepSeek model led on the Artificial Analysis index and on several coding measures. On Terminal Bench, Inkling 2 Small scored 13.1% against 33.8 for the DeepSeek model. The reviewer still judged DeepSeek the stronger model overall. Inkling 2 Small pulled ahead on vision, knowledge work, and finance, with top scores on several of those tests. In hands-on checks it generated a self-contained HTML animation from one prompt, produced a multilingual proverb table covering many scripts, and turned a hand-drawn circuit image into a working simulator with error analysis. In a separate vision example, it identified the grape and vintage in three glasses of red wine from a photo, correct in 19 of 20 trials. The lab lists its own limits: it is not the most concise communicator, has no audio support, and is less polished as a chat model. If your workload is plain back-and-forth chat, a premium-priced model like this is a poor fit. ## Frequently Asked Questions ### Can I run Inkling 2 Small locally? No. At the time of the review it was available only through the Thinking Machines console. The reviewer hoped for a permissive open release later, but nothing was announced. ### Do I need to pay before I can try it? Yes, based on the reviewer's experience. They had to get an API key and add credits to the account before testing. Check the console for any free allowance or trial terms. ### What are the per-token prices? The material behind this article does not state them. The reviewer only described the model as expensive. Look up current input and output rates in the console before you budget a project. ### Does Inkling 2 Small accept images? Yes. It reads text and images, and the reviewer sent an image through the API. It does not support audio. ### Where does the $426 figure come from? It comes from a lab demo in which the model autonomously fine-tuned a small forecasting model over about four hours. It is one project's total cost, not a standard rate, and the breakdown between model tokens and training compute was not given.