What changed
Thinking Machines Lab released Inkling 2 Mini under Apache 2.0 with 26.4B total parameters, 3.4B active parameters per pass, 5.4B effective parameters, multimodal input and a 1M-token context. The looped architecture can vary compute at inference and already has support in Transformers, llama.cpp, MLX and SGLang.
Why it matters
A compact open-weight model with variable-pass reasoning gives local operators a new cost-versus-depth control and provides researchers a public artifact for testing looped architectures.
What to do
Pin the model and runtime, start on a 32GB machine, cap passes and tokens, then compare task success, latency, memory, looping failure, tool safety and multilingual quality against a fixed-compute local baseline.
Caveat
Published benchmark results are vendor evidence. The model card warns about hallucinations, unreliable or destructive tool use, repetition, weaker non-English behavior and loops that do not map cleanly to lower compute.