Inkling 2 Mini drops for local coding
Thinking Machines Lab released Inkling 2 Mini, an open-weight 26.4B Mixture-of-Experts multimodal model with 5.4B active parameters and a 1M-token context window. Built with a looped transformer architecture, it runs locally on consumer laptops while scoring 48.2% on DeepSWE v1.1 and 77.9% on SWE-bench Pro v2.
Inkling 2 Mini proves that looped transformer architectures can squeeze frontier-level coding and agentic performance into consumer-grade hardware. Local developer tooling no longer requires compromising on model depth or context length.
- –Middle 12 layers loop three times to emulate a 48-layer network while maintaining a compact 26.4B footprint and 5.4B active parameters per token.
- –Scores of 48.2% on DeepSWE v1.1 and 77.9% on SWE-bench Pro v2 outperform much larger open models on long-horizon software engineering tasks.
- –A 1M context window and native image understanding enable full-codebase reasoning and multimodal agent debugging on a 32GB laptop.
- –Apache 2.0 weights with immediate support in llama.cpp, MLX, and Hugging Face make it a drop-in driver for local-first AI coding assistants.
DISCOVERED
17m ago
2026-11-20
PUBLISHED
35m ago
2026-11-20
RELEVANCE
AUTHOR
0x0SojalSec