← Back to blog
#IA#Modelos abiertos#Programación#Estudiantes#Guía

Inkling 2 Small and Open Models: A Checklist for Developers and Students Before Adopting One

Colorful Python code on a dark screen

This week Thinking Machines, Mira Murati's company, introduced Inkling 2 Small, a model that aims to be the best "open-weight" model of Western origin. According to its official announcement and coverage from The Decoder, it's a mixture-of-experts (MoE) model with 594 billion parameters in total and 17 billion active per token, which accepts text and images and is designed as a base for fine-tuning.

Beyond the news, it's a good excuse for a Saturday study session: how to evaluate an open model before putting it in your software, whether you code, study or lead a technical team.

What we know about Inkling 2 Small

  • Architecture: MoE with 594 billion total and 17 billion active parameters. It accepts images and lets you adjust reasoning effort.
  • Availability at launch: as a proprietary service on the Tinker platform. The weights were to be released under an Apache 2.0 license once external safety testing was complete, according to the announcement.
  • Results: the company and The Decoder report large improvements over its previous version on agentic coding tests. These are figures from the maker and from external evaluations cited by the press: treat them as an indication, not a guarantee for your case.
  • Limitations the company states: it can be wordy, doesn't always pick the best tool in large codebases, doesn't support audio and is less optimized for chat.

Update (November 18, 2026): the official page says the weights were published on November 12 on Hugging Face, after safety testing was completed.

Why "open" doesn't always mean the same thing

Three different things are often mixed up: downloadable weights, a license that allows commercial use and open code and training data. A model can have the first without the second. Before building a product on it, read the full license, not the headline. For an overview, see our article on open-source AI models.

What this means for your SME or technical team

An open model can be attractive when you need control over your data (so it doesn't leave your infrastructure), predictable costs or customization for a specific sector. But it also brings work: someone has to deploy, monitor, update and secure it. At 594 billion parameters, Inkling 2 Small is a large model: you'd probably use it through a provider that hosts it, not on an office server.

A checklist to evaluate a model before adopting it

  1. Define the task with real examples. Gather 30 to 50 cases from your business (customer queries, code snippets, documents) with the expected answer. A public benchmark doesn't measure your problem.
  2. Compare against what you already use. Test it side by side with your current model, blind if possible, and measure quality, response time and cost per task.
  3. Measure total cost, not price per token: a model that "thinks" more generates more tokens per task and can end up costing more.
  4. Check the license: commercial use, user limits, sector restrictions, attribution requirements.
  5. Ask where your data runs. If you use a provider, what it stores, for how long and whether it trains on your inputs. Keep in mind the SPDP rule on AI and personal data.
  6. Test failures. Ask for impossible, ambiguous or false-premise things. Measure how much it makes up and whether it admits it doesn't know.
  7. Plan your exit. Make your application able to switch models without a rewrite: a simple abstraction layer and automated regression tests.
  8. Pin versions. Use a specific model version and re-evaluate when it changes.

If you're just starting, our guides to tools for developers and free resources for students are a good place to practice with local models and code assistants.

A warning about benchmarks

Results published by the maker are usually measured with configurations it chose. Until there are independent evaluations and the community can reproduce them with the published weights, treat them as hypotheses. And remember that a model can top a coding benchmark and still fail on your code, your language or your domain.

How We Approach It at SimCodec

At SimCodec we build software and AI integrations and help choose the right model for each case: tests with real data, cost control and an architecture that lets you switch providers. If you're evaluating an AI solution for your company, write to us or call our AI assistant Cyntia at +593 99 726 6838.

Sources

← Back to blog Get a quote →