Thinking Machines Lab released Inkling 2 Mini, an open-weights 26.4-billion-parameter looped transformer that runs its middle layers three times per token for better reasoning with fewer tokens on agentic coding and similar tasks1,4. Artificial Analysis rated the model at 20 on its Intelligence Index, matching Qwen3.8 27B (xhigh) with roughly five times fewer active parameters2.
Architecture and footprint
The model is a Mixture-of-Experts architecture with 3.4 billion active parameters and 5.4 billion weight activations per token once looping is accounted for. Its 24 layers are split into four blocks of six layers; the first and last blocks run once, while the two middle blocks run three times, yielding 48 layer applications from 24 stored layers3. Its total memory footprint is roughly 20 GB, small enough to run on a consumer laptop.
Looping allows the model to reason longer without writing its reasoning into the chain of thought, which TML says matters for monitoring: the ability to oversee and understand what the model does by reading its chain of thought.
Benchmark results
TML reports that Inkling 2 Mini leads every open-weights model under 100 billion total parameters on DeepSWE v1.1 at 48.2% and Vals Index v2.2 at 46.5%. Additional scores reported by TML include 79.2% on SWE-bench Verified, 77.9% on SWE-bench Pro v2 Hard, 73.7% on SWE-bench Multilingual, 71.6% on Terminal Bench v2.1, 77.9% on Vibe Code Bench v1.1, 50.0% on MultiChallenge, 82.7% on MCP Atlas, 85.3% on DeepSearchQA, and 79.0% on BrowseComp. On PMPP-Hard, Inkling 2 Mini scored 455.82, ahead of Qwen3.8-27B at 262.50. TML also reported 0.703 intelligence per joule on an NVIDIA DGX Spark. TML says the model runs faster than similarly performing models.
TML also released TML-d24-varloop-research, a variable-loop variant, alongside the main model. The variable-loop model's no-chain-of-thought time horizon improved from 16 seconds at one pass to 37 seconds with three loops, and its performance on a serial-reasoning benchmark kept improving out to ten loops, twice as many as it was trained with. TML noted the model is not the strongest small model for multi-turn chat or for instructions with many constraints.
ANALYSIS A model that reaches answers with fewer visible reasoning tokens narrows the surface an auditor must review, making the monitorability framing a double-edged claim: less CoT text is easier to read but also offers less insight into the model's internal process.
Inkling 2 Mini is available on Tinker for inference and customization with open weights, and full weights are downloadable from Hugging Face. The release follows TML's Inkling 2 Small, which was pretrained from scratch in 68 hours on 12 trillion tokens using native NVFP4[4], and comes after reports that Accel is leading a $5 billion round for TML at a $40 billion pre-money valuation[5].