Inkling 2 Mini ties Opus on CyberBench
Thinking Machines Lab released Inkling 2 Mini, an open-weights 26.4B looped-transformer model that ties Claude Opus 5.5 at 74.58% on Vals AI's CyberBench. The model runs locally with zero refusal guardrails after community ablation, bringing frontier cybersecurity and offensive reasoning capabilities to consumer laptops.
An open-weights model matching frontier closed models on offensive security while running locally on consumer hardware is a watershed moment for AI security research. The instant stripping of refusal layers underscores how difficult post-training safety guardrails are to enforce once weights are public.
- –Looped transformer architecture: With 26.4B total parameters but only ~3.4B active per token across recurrent decoder layers, Inkling 2 Mini fits into roughly 20GB of RAM for local inference on 32GB MacBooks.
- –Benchmark parity: Scoring 74.58% on Vals AI's CyberBench puts it in a dead heat with Claude Opus 5.5 (CVP), proving compact architectural designs can rival frontier models in vulnerability exploitation.
- –Zero-refusal ablation: Community abliteration dropped refusal rates on harmful prompts from 81/81 to zero, removing safety filters without degrading exploit generation or code-synthesis performance.
- –Offensive accessibility: Developers and penetration testers gain an offline, zero-cost automated red-teaming agent without cloud API telemetry or restrictive corporate use policies.
DISCOVERED
22m ago
2026-11-22
PUBLISHED
50m ago
2026-11-22
RELEVANCE
AUTHOR
0xCarnagee