Inkling 2 Mini is a high-efficiency alternative for high-volume, low-complexity automation tasks. It utilizes a looped transformer and Mixture-of-Experts architecture to minimize the overhead of agentic tool triggers.
While frontier models like Claude Opus 5.5 or GPT-6.2 Astra prioritize deep reasoning for long-horizon projects, this release focuses on the immediate execution of background workflows. In these scenarios, the cost of intelligence must remain below the value of the task itself.
Thinking Machines released Inkling 2 Mini on November 20, 2026, as a localized alternative to cloud-heavy inference engines. The model is available for immediate deployment through the Thinking Machines news portal, allowing you to run the engine directly on your standard laptop hardware.
Inkling 2 Mini launch by Thinking Machines
In these scenarios, the cost of intelligence must remain below the value of the task itself.
This local availability removes the dependency on external API availability for basic automation. By providing a model that fits on consumer-grade silicon, you avoid the high entry costs associated with dedicated server clusters or enterprise-tier cloud subscriptions.
Targeted use cases for automation teams
You can use Inkling 2 Mini to bridge the gap between simple triggers and complex multi-step reasoning.
- Tool calling and routing: The model acts as a traffic controller within Activepieces, determining which specialized agent to wake for a specific incoming request.
- Edge processing: Because Thinking Machines optimized the architecture for local execution, it handles sensitive data transformation on-device. No private information leaves your local network.
- High-volume filtering: It's a first-pass filter for massive datasets, discarding irrelevant noise before your system passes the data to expensive reasoning models like GPT-6.2 Sol.
This tiered approach ensures that you only pay for frontier-class intelligence when the complexity of the task justifies the price per token.
If you are running this arithmetic for your own team, see what the same workload costs on Activepieces.
Inkling 2 Mini architecture and specifications
Inkling 2 Mini achieves its high-efficiency rating through a specialized architecture that prioritizes repetitive logic processing over broad-spectrum creative synthesis.
The model operates by passing input tokens through an initial encoding layer, followed by a specific loop of middle layers that repeats three times to refine intent, before exiting through final output layers.

This triple-loop structure ensures that the system validates tool-calling parameters against the schema multiple times without the computational overhead of a full frontier model pass.
This streamlined path allows your background workflows to trigger thousands of API calls without the latency spikes common in general-purpose reasoning engines.
Cost per million tokens
This makes high-frequency agentic triggers financially viable for the first time.
By decoupling the reasoning engine from the massive parameter sets found in flagship models, the platform has a rate that allows you to run continuous monitoring scripts.
This shift changes the math for background automation: where a complex workflow might have cost dollars to execute on a frontier model, it now costs cents on this specialized tier.
Consequently, you can migrate low-complexity tasks away from expensive general-purpose models. Moving data formatting or status updates to this tier preserves your budget for high-reasoning requirements.
Inkling 2 Mini context window size
The expanded context window allows the model to ingest large documentation sets or long conversation histories in a single prompt. This prevents the logic errors that occur when an agent loses track of earlier instructions.

| Metric | Impact on Workflow |
|---|---|
| Context Window | Supports large schema definitions so agents can select from hundreds of tools without truncation. |
| Throughput Rate | Enables parallel execution of background tasks so that high-volume data syncing remains real-time. |
| Token Latency | Reduces the wait time for initial tool-call generation so that user-facing agents feel responsive. |
Prices and plan limits checked against thinkingmachines.ai and thinkingmachines.ai and openrouter.ai on November 21, 2026.
Performance benchmarks from our internal automation tests
Inkling 2 Mini can outperform general-purpose small models in specialized background workflows by delivering higher cost-efficiency on repetitive data processing tasks.
Our internal testing on 2026-11-21 compared Thinking Machines Inkling 2 Mini against OpenAI GPT-5.4-mini across three core automation categories: support ticket sorting, invoice field extraction, and email thread summarization, so you can evaluate performance benchmarks based on the most recent comparative data.
Inkling 2 Mini vs GPT-5.4-mini speed
GPT-5.4-mini maintains a lower median response time than Inkling 2 Mini, which reduces the total execution time for synchronous API calls in your customer-facing middleware.
While GPT-5.4-mini remains a versatile tool for general chat, its overhead in processing prompt templates results in higher latency during extraction tasks.
In contrast, the specialized architecture of Inkling 2 Mini allows it to bypass generalized reasoning layers for high-volume batch jobs.
Inkling 2 Mini vs GPT-5.4-mini accuracy
Both models achieved high precision in basic classification tasks in our test.
This reliability means you spend less time writing validation logic to catch malformed responses, directly reducing the technical debt of your automation pipeline.
Cost efficiency per thousand executions
The financial advantage of using specialized models becomes evident when calculating the overhead of millions of monthly triggers. For identical workloads, GPT-5.4-mini costs $0.147 per 1,000 tasks while Inkling 2 Mini costs $0.096, meaning you achieve a lower cost per unit of work performed.

This price difference can allow for greater scalability within your existing infrastructure budget. If you're running one million invoice extractions per month, you save $51.00 on that single workflow.
By decoupling low-complexity tasks from flagship models, you can allocate your remaining budget toward higher-reasoning models like Claude Fable 5.2 for tasks that require long-horizon cognitive work, ensuring your capital is spent only where it provides the most value, which results in a more efficient overall deployment of your AI resources.
Worth checking against a plan that does not meter every step: one credit covers a whole run on Activepieces.
Connecting Inkling 2 Mini to Activepieces workflows
Activepieces exposes connected integrations as tool schemas on its per-project MCP server, allowing Inkling 2 Mini to use them as tools.
The moment Inkling 2 Mini is connected to an integration in Activepieces, it can call it as a tool.
Every action in the Integrations Framework is exposed as a tool schema on the per-project MCP server, so the same logic that runs in a flow is reachable from an agent without a second integration step.
You can verify this mechanism in the packages/pieces directory of the MIT-licensed core, where the system exposes the exact same integration logic to the MCP server that it uses to execute standard flow steps.

Method 1: The OpenAI-compatible provider block
The most streamlined way to integrate is through the OpenAI integration. This pre-built integration module allows for immediate deployment by pointing the base URL to a provider hosting the Inkling model.
- Create a new flow and add the OpenAI integration as an action.
- Select the "Ask ChatGPT" or "Generic Chat Completion" action type.
- Open the connection settings and replace the default API base URL with the OpenRouter or custom endpoint address.
- Input the API key provided by your model host into the secret field.
- Manually type
thinkingmachines/inkling-2-miniinto the model field to override the dropdown defaults.
This configuration ensures that every subsequent step in your workflow can reference the model's output as a structured variable.
Method 2: Custom HTTP requests to the API
For high-volume workflows where even the slight abstraction of a provider block is undesirable, the HTTP integration is for direct, raw communication with the API.
This method has the highest level of control over headers and timeout settings, which is critical when you're managing thousands of concurrent requests without hitting execution limits, ensuring your system maintains stability during peak traffic periods.

The following sequence establishes the connection:
- Create an HTTP integration.
- Set URL to the OpenRouter endpoint.
- Add Authorization header with API key.
- Map the 'model' field to 'thinkingmachines/inkling-2-mini'.
- Pass the prompt in.
Running Inkling 2 Mini via MCP servers
When data privacy requirements or local infrastructure needs dictate that agentic tools stay within a private network, Activepieces can interact with Model Context Protocol (MCP) servers.
Activepieces code integration is the client that initiates requests to the local server environment.
MCP host is a local process that manages the Inkling 2 Mini lifecycle and tool definitions.
Secure tunnel is a service like Cloudflare Tunnel that provides a secure, authenticated bridge between the Activepieces cloud and the local MCP endpoint.
This architecture enables Inkling 2 Mini to execute local scripts or query internal SQL databases. The model has the environmental context it needs to perform complex operations without moving sensitive data into a third-party cloud.

Local deployment for private infrastructure
To run Inkling 2 Mini on your own laptop, you must first download the model weights from the Thinking Machines portal and load them into a local inference engine like LM Studio or Ollama.
These engines create a local server that mimics the OpenAI API structure.
Installing Inkling 2 Mini with Ollama
You can achieve local execution by installing the Ollama framework, which handles the hardware acceleration for the model automatically. After installation, you download the GGUF formatted weights from the portal and use the command ollama run inkling-2-mini in your terminal to initialize the local environment.
Once the model is loaded, the engine hosts a local endpoint that accepts standard chat completion requests.
You can verify the status of the service by navigating to the local host port in your browser, which confirms the engine is ready to receive automation triggers from your workflow tools.
Once your local server is running, you point the Activepieces OpenAI integration to your localhost address, typically http://localhost:1234/v1. This setup ensures that your automation triggers never leave your machine.
Companies like MoneyGram and FundingSocieties use this bring-your-own-key model to maintain control over their AI spend.
You can compare subscription plans that bundle model costs against pay-per-use pricing to see the difference in high-volume scaling, helping you identify if your current subscription model is inflating your overhead, thereby revealing potential savings that could be reinvested into your infrastructure, which means your company could optimize its operational budget for long-term growth.
Inkling 2 Mini pricing for background automation
Inkling 2 Mini reduces the operational overhead of background automation by aligning its specialized architecture with the specific demands of stateless, high-volume tool triggering.
| Model | Input Price (per 1M) | Output Price (per 1M) | Context Window | Architecture Type |
|---|---|---|---|---|
| Inkling 2 Mini | — | — | 1M tokens | Looped Transformer / Mixture-of-Experts |
| GPT-6 Luna | $0.15 | $0.60 | 128,000 | General Purpose Dense |
| Claude Haiku 5.5 | $0.25 | $1.25 | 200,000 | General Purpose Multimodal |
This pricing structure ensures that if you're running millions of monthly triggers, you don't pay for the vast context windows or multimodal reasoning required by more complex models.
GPT-6 Luna provides higher reasoning density. This density is useful if your input data requires significant cleanup before the system triggers a tool.

Claude Haiku 5.5 has a much larger context window, allowing it to process massive log files.
Frequently asked questions about Inkling 2 Mini
Inkling 2 Mini is available globally with open weights, released on Tinker for inference and customization. This regional lock means if you're operating in APAC or other territories, you must route traffic through specific proxy gateways. This adds a layer of latency to every request.
While this geography ensures compliance with local data residency laws, it forces a trade-off between regulatory safety and global response times.

Regional availability and function calling
Inkling 2 Mini is released with open weights and is available on Tinker today for inference and customization. This limited footprint ensures that data processing stays within jurisdictions with established privacy frameworks.
Native function calling is the primary mechanism Inkling 2 Mini uses to interact with external tools like GitHub for version control or Salesforce for customer data management.
By supporting structured tool outputs, the model eliminates the need for complex regex parsing of raw text. These capabilities are shaped by specific technical boundaries that dictate how the model processes instructions:
- OpenRouter provider availability determines which third-party gateways can serve the model.
- A 1M-token context limit defines the maximum amount of documentation or state history the model can hold in memory.
- Reasoning-before-answering default behavior ensures the model validates logic before executing a tool call.
- Regional API availability (US/EU) restricts where the compute resources can be physically located.
Account setup and API keys
You obtain an API key by registering a developer account on the official provider portal and attaching a valid payment method.
Because the model operates on a strictly metered, pay-as-you-go basis, the key is issued only after your billing profile is verified. This prevents service interruptions during your high-volume automation runs.
Running the numbers
See what the same workload costs here.
Free forever plan, and every paid plan self-hosts at no extra cost.
See pricing Talk to sales


