Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, offering an Apache 2.0-licensed, multimodal model designed for agentic applications and local execution. It has 26.4 billion total parameters, 3.4 billion active parameters, and a roughly 20GB memory footprint, with support for Ollama, llama.cpp and several server-oriented inference frameworks. The model card lists Apple silicon devices with at least 32GB of memory among its supported hardware.
The architectural detail worth noticing sits behind those parameter counts. Inkling 2 Mini repeatedly processes each token through its middle layers, using the same weights multiple times. That increases its effective computational depth without requiring a separate set of weights for every layer application.
For developers, the release offers two things worth separating: a permissively licensed model that can fit on some consumer machines, and vendor-reported coding results that make it a candidate for local agents. Its small active-parameter count alone, however, is not enough to predict speed or memory use.
How Inkling 2 Mini’s Looped Transformer Works
Inkling 2 Mini is a 24-layer, decoder-only transformer with sparse mixture-of-experts feed-forward layers. For each token, each layer routes work to eight of 128 experts, according to the architecture description.
The middle 12 layers run three times with shared weights. The arithmetic is straightforward:
- The 12 layers outside the loop run once.
- The middle 12 layers run three times, producing 36 layer applications.
- Together, that gives 48 layer applications from 24 layers of weights.
Shared weights save storage compared with giving every application its own distinct weights. They do not make the extra processing free.
That explains why the release carries three different parameter figures:
| Figure | What it describes |
|---|---|
| 26.4B total parameters | The model’s total parameter count |
| 3.4B active parameters | The active count before accounting for repeated loop passes |
| 5.4B effective active parameters per token | The count after including the extra applications of the looped layers |
Each additional pass through the middle layers adds about one billion effective active parameters. Two extra passes take the count from 3.4B to 5.4B.
My view is that 5.4B is the more useful figure when interpreting the model’s computational workload. The 3.4B number remains a valid architectural specification, but it leaves out repeated work. Neither figure, by itself, predicts a laptop’s tokens-per-second rate.
Thinking Machines describes looping as a way to “reason longer without writing its reasoning down.” That means additional processing happens inside the network, without necessarily producing more visible reasoning tokens. The announcement presents this as particularly useful for reasoning-heavy agentic coding.
What the Roughly 20GB Footprint Includes
Thinking Machines distributes Inkling 2 Mini as a single checkpoint of about 20GB, using mixed numerical precision. Most weights were trained and deployed natively in NVFP4, with the final four layers using MXFP8, according to the release announcement.
Those low-precision formats are an important part of the local-execution story. A parameter count describes how many learned values exist; the numerical format determines how much storage those values require.
The model also supports:
- Text and image inputs, with text output.
- A context window of up to one million tokens.
- Images encoded through a vision transformer and processed jointly with text.
- English, broader multilingual use and multiple programming languages.
The model card recommends image dimensions between 40 and 4,096 pixels for optimal performance.
A roughly 20GB checkpoint should not be read as a promise that the entire application needs exactly 20GB. Inference software, working buffers, context caches, the operating system and other applications also need memory.
The practical consequence is especially important for that million-token context limit: fitting the weights and running the maximum context are separate requirements. I would treat the published context window as a capability ceiling, then choose a working context length that fits the machine and workload.
How to Run Inkling 2 Mini Locally
The model card lists local deployment paths for Apple silicon, NVIDIA systems and GGUF-based inference. The right starting point depends on both the hardware and the checkpoint format.
| Hardware or deployment route | Listed software support | Hardware guidance |
|---|---|---|
| Apple silicon | Ollama, Exo Labs | 32GB or more unified memory |
| NVIDIA systems | SGLang, vLLM, Unsloth, NVIDIA Dynamo, Tokenspeed | One NVIDIA DGX Spark is explicitly supported |
| Other GPUs and platforms | Depends on the framework | 32GB or more VRAM recommended |
| Quantized GGUF deployment | Unsloth and llama.cpp | Requirements depend on the selected quantization and configuration |
These are the routes documented in the model card, not equivalent performance guarantees across every framework.
Apple Silicon: Start With Ollama or Exo Labs
MacBook, Mac mini and Mac Studio owners with at least 32GB of memory fall within the listed Apple silicon configuration.
A sensible setup sequence is:
- Open the Ollama or Exo Labs integration linked from the model card.
- Install a version that supports Inkling 2 Mini.
- Obtain the model package specified by that integration.
- Start with a modest context setting and a simple text request.
- Connect an agent client after basic inference works.
Use the integration’s current model identifier and launch instructions. A generic command copied from another LLM’s setup guide may select the wrong package or assume an unsupported architecture.
llama.cpp: Choose the GGUF Distribution
The model card lists quantized GGUF versions through Unsloth and llama.cpp. This is the documented route for readers using a GGUF-based local setup.
Choose the quantization your hardware can accommodate, and check the release-linked instructions for the required runtime build and launch settings. Different quantizations change storage and can affect output quality, so the smallest download is not automatically the best choice for a coding agent.
For server deployments, SGLang and vLLM are among the listed NVIDIA options. The distributed mixed-precision checkpoint and a GGUF conversion are different deployment packages; follow the instructions for the one you actually download.
Add an Agent Layer Separately
A model runtime generates responses. An agent application supplies the surrounding loop: presenting tools, executing approved calls, returning results and deciding when work is complete.
That distinction matters when reading claims about local agentic work. Installing the weights is one step; configuring shell access, repository access and tool execution is another.
I would begin with a disposable repository and constrained file and shell permissions. Local inference keeps model computation on the device, but an agent using web search or networked tools can still send information outside it.
What the Vendor Benchmarks Actually Show
Thinking Machines says Inkling 2 Mini leads open-weight models below 100B total parameters on benchmarks including DeepSWE v1.1 and Vals Index v2.2. Its published evaluation table includes:
| Benchmark | Inkling 2 Mini result | Published setup detail |
|---|---|---|
| DeepSWE v1.1 | 48.2% | mini-swe-agent v2.4.6 |
| SWE-bench Pro v2 Hard | 77.9% | mini-swe-agent v2.4.6; announcement identifies a 51-task hard subset |
| Vals Index v2.2 | 46.5% | Reported index score |
These are vendor-published results, not independent measurements supplied with this coverage. The announcement also states that evaluation setups vary by model, which limits how cleanly every row can be compared.
Frequently Asked Questions
3 questions
1What License Does Inkling 2 Mini Use?
The model card lists Apache 2.0. Its open weights are intended to support research, fine-tuning and integration into third-party products.
