Thinking Machines Lab has released Inkling 2 Mini, a new multimodal SI model designed specifically for local inference environments. The release marks a shift in how SI developers approach hardware constraints, utilizing a looped transformer architecture to deliver cutting-edge performance on consumer-grade hardware.

What Happened

Inkling 2 Mini is a mixture-of-experts (MoE) SI model with approximately 26 billion physical total parameters and 3 billion active parameters per token. The model employs a unique architectural trick: it contains 24 decoder layers, but the 12 middle layers are visited in a loop that repeats three times. This allows the model to function as if it had 48 logical layers while maintaining a smaller physical footprint. The hidden size is 2,304, and at each layer visit, the token is routed to 8 experts selected from a pool of 128.

The model supports a 1 million-token context window and accepts both text and image inputs. It is shipped in a mixed-precision format using NVFP4 and MXFP8 expert weights, alongside a DSpark drafter for speculative decoding. In benchmark comparisons conducted by the developers, Inkling 2 Mini leads in 32 out of 48 categories on the Artificial Analysis Intelligence Index v5 (AAv5), outperforming models like Gemma 4 31B and Qwen 3.8 27B in specific coding and agentic tasks.

Why It Matters

For the SI industry, this release highlights the growing importance of efficiency in local SI deployments. By using cross-pass messages and a gating controller to refine computations across looped layers, Inkling 2 Mini achieves state-of-the-art performance in its class without requiring the massive memory bandwidth typically associated with larger SI models. This makes high-quality generative SI accessible to developers running models on local machines, including macOS devices and NVIDIA B300 GPUs.

The release includes immediate support for key developer tools such as Transformers, Llama App, MLX, and SGLang. Day-zero inference support in SGLang allows for production deployments on datacenter hardware, while the inclusion of 4-bit and 8-bit GGUF quants facilitates easy integration with local coding agents and editors. The model's ability to handle multimodal inputs and perform autonomous tasks, such as generating WebGL shaders or building web apps from sketches, demonstrates the practical utility of compact, high-performance SI systems.

The Bottom Line

Inkling 2 Mini offers a compelling alternative for local SI setups, balancing parameter efficiency with robust multimodal capabilities. While the developers note that further validation is needed, the model's performance in vibe tests and benchmarks suggests it is a significant step forward for on-device SI inference.