Uncensored Inkling-2-Mini (26B MoE) runs locally on 16GB, scores 77.9% on SWE-bench Pro Hard
udmrzn · x · 2026-11-22
A community-released uncensored build of Inkling 2 Mini can now run locally: a 5.4B-active-param MoE with a looped transformer architecture, 1M context support, and it fits in 16GB of memory.
Key numbers:
- 48.2% on DeepSWE v1.1
- 77.9% on SWE-bench Pro (Hard), claimed best-in-class for its size on coding/agentic work
- Uncensoring was done by removing the refusal direction from the residual stream
- GGUF builds are ready for Unsloth + llama.cpp, with LoRA adapters on the way
Related event: Inkling 2 Mini local tests: strong coding, uneven deployment experience(8 posts)→
More from coding & agent
- fofrAI breaks down the workflow behind his vibe-driven AI images, with copyable agent instructions — fofrAI · 2026-11-22
- Agentic Benchmark: Gemini 4 Hydrogen Costs $0.37 Per Task vs Neon's $0.84 — Money-String4165 · 2026-11-22
- Dev sends 7 PRs to open-source coding agent Shelley after switching to Gemini — swaroopch · 2026-11-22
- Databricks previews agent platform that lets you git-branch your entire agent — matei_zaharia · 2026-11-22
- Taste MCP reins in AI's tendency to bloat UI with unneeded text and elements — annetgriffin · 2026-11-22
- Daniel Lemire: why you should write your own pull-request messages — lemire · 2026-11-22