Inkling 2 Mini tops Artificial Analysis index
Thinking Machines Lab's open-weight 26B model outscored Claude Haiku 5.5, MiniMax-M3, and GPT-6 Luna Low on the Artificial Analysis Intelligence Index. Its looped transformer architecture keeps active parameters down to 3.4B, enabling high-performance agentic and coding workflows locally on 16GB laptops.
Looped architectures are proving parameter count is no longer a proxy for reasoning power, shifting the frontier race directly to developer laptops. Thinking Machines Lab is demonstrating that recurrent layer re-use can rival proprietary small models without requiring massive cloud infrastructure.
- –Looped execution visits middle layers three times per token, effectively mimicking a 48-layer network while maintaining a 3.4B active parameter footprint.
- –Beating hosted offerings like Claude Haiku 5.5 and GPT-6 Luna Low eliminates the performance penalty typically expected from running local models offline.
- –A 1-million-token context window paired with sub-4B active inference enables viable autonomous code loops and tool-calling on 16GB laptops.
- –Apache 2.0 licensing and immediate compatibility across Ollama, MLX, and vLLM make it an immediate target for local agent stacks.
DISCOVERED
7m ago
2026-11-21
PUBLISHED
36m ago
2026-11-21
RELEVANCE
AUTHOR
0x0SojalSec