Thinking Machines Lab launches Inkling 2 Mini, an open-weight 26.4B parameter multimodal model featuring a unique looped transformer architecture for local edge compute.

On November 20, 2026, Thinking Machines Lab expanded its high-performance Inkling family with the release of Inkling 2 Mini. While the recently released Inkling 2 Small focused on enterprise-grade frontier performance with a massive 594B parameter MoE, the Mini variant is precision-engineered for a different mission: high-efficiency, local-first multimodal intelligence.
Released under the Apache 2.0 license, Inkling 2 Mini represents a significant shift toward making sophisticated multimodal reasoning accessible on consumer-grade hardware. By combining open weights with an innovative architectural approach, Thinking Machines Lab is providing developers with a model that bridges the gap between massive cloud-based LLMs and the practical constraints of edge computing.
The technical core of Inkling 2 Mini is its 'looped transformer' architecture. Unlike standard decoders, this 24-layer model features a specialized middle section where the middle 12 layers run three times using shared weights. This design allows the model to perform 48 layers of applications per token while only maintaining the weight footprint of 24 layers, effectively increasing computational depth without a linear increase in VRAM requirements.
The model utilizes a Sparse Mixture-of-Experts (MoE) strategy, totaling 26.4B parameters but only activating 3.4B per token (8 of 128 experts per layer). Because of the looping mechanism, each extra pass adds approximately 1.0B effective active parameters, resulting in a total of 5.4B effective active parameters per token. This provides the reasoning density of a much larger model with the latency of a small one.
Multimodality is handled natively through a vision transformer (ViT) encoder that projects both text and image inputs into a shared hidden space. This allows for seamless cross-modal reasoning rather than relying on a separate projection layer. Furthermore, the model supports a massive 1M token context window, supported by a hybrid local and global attention mechanism.
Inkling 2 Mini is strategically positioned for on-device and edge use rather than competing directly with the 27B+ parameter heavyweights. In comparative testing, the model shows remarkable efficiency, achieving a Third Party Index score of 4.4. It is designed to punch above its weight class, offering reasoning capabilities that rival models like Gemma 4 31B and Qwen3.8-27B in specific multimodal tasks.
Benchmark comparisons show that while it does not attempt to match the absolute frontier scores of the 594B Inkling 2 Small, it outperforms many contemporary models in the sub-30B category, including Muse Glimmer and Nemotron 3.5 Super VL, particularly in tasks involving complex image-to-text reasoning and long-context retrieval.
The model was evaluated using a standardized suite of multimodal reasoning and instruction-following benchmarks. The evaluation methodology focused on 'intelligence-per-watt' and 'intelligence-per-GB,' ensuring that the performance metrics reflected real-world utility for developers running models on local hardware like MacBooks or NVIDIA DGX Spark systems.
For developers who prefer not to host the model locally, Inkling 2 Mini is available via OpenRouter. The pricing is highly competitive, sitting in the same tier as the GPT-6 Luna class models, making it an extremely cost-effective choice for high-volume multimodal applications.
This pricing structure allows for rapid prototyping and scaling without the heavy overhead typically associated with multimodal API calls. The cost per million tokens is optimized for both high-frequency chat interfaces and heavy-duty document/image processing pipelines.
Due to its 20 GB checkpoint size and optimized mixed-precision (NVFP4 and MXFP8), Inkling 2 Mini is a prime candidate for local AI agents. It is particularly well-suited for privacy-sensitive applications where data cannot leave the device, such as personal assistants running on Apple Silicon (MacBook/Mac Studio) or local enterprise edge nodes.
Developers can leverage its 1M token context window for advanced RAG (Retrieval-Augmented Generation) tasks, allowing the model to 'read' entire technical manuals or 'view' long sequences of visual data. Its multimodal nature also makes it ideal for coding assistants that can interpret UI screenshots, or for automated visual inspection in industrial edge computing environments.
Inkling 2 Mini is ready for immediate deployment. The weights have already been integrated by the GGUF quantization community, making it compatible with popular local inference engines like llama.cpp and Ollama. For high-throughput serving, support is available via vLLM, SGLang, and NVIDIA Dynamo.
For developers looking to customize the model, Thinking Machines Lab provides the Tinker platform, which includes a playground and a comprehensive cookbook for fine-tuning. Post-training can be further optimized using the Miles and SkyRL frameworks to align the model with specific agentic or conversational behaviors.
API Pricing β Input: 0.10 USD per million tokens (OpenRouter) / Output: 0.40 USD per million tokens (OpenRouter) / Context: 1M token context window