Inkling 2 Mini: A 20GB Open Coder That Thinks in Loops, With a Check on What It Hides
A model that writes code at near frontier-small level now fits on a laptop, and the lab that built it also measured whether it thinks things it never writes down. Thinking Machines released Inkling 2 Mini on Friday with open weights under Apache 2.0 on Hugging Face. It is a Mixture-of-Experts model with 26.4B total parameters and 3.4B active. The middle two of its four blocks run three times for every token, so 24 stored layers become 48 layer applications and 5.4B weight activations per token. It is trained natively in NVFP4, with the last four layers in MXFP8, so the roughly 20GB footprint carries no quantization loss. The team ran it offline on a 48GB M5 Pro MacBook in airplane mode and had it build a flight simulator. Post: https://thinkingmachines.ai/news/inkling-2-mini/ Weights: https://huggingface.co/thinkingmachines/Inkling-2-Mini
The reported numbers are strong for its size. DeepSWE v1.1 48.2% against 42.2% for Qwen3.8-27B, SWE-bench Verified 79.2% bash-only, SWE-bench Pro v2 Hard 77.9%, BrowseComp 79.0%, MCP Atlas 82.7%. It does not win everywhere. It trails Qwen3.8-27B on Terminal Bench v2.1 (71.6% vs 73.0%), ties it on the Artificial Analysis index at 20.1, and is weaker at multi-turn chat and many-constraint instructions. Looping adds reasoning, not knowledge, so factual recall tracks its size, and it makes up the gap with search and tools. All of these are vendor-reported.
The bigger story is monitorability. Chain-of-thought monitoring works because a normal transformer can only do so much computation per token before it has to write something down. Looping raises that limit, which is exactly why safety people worry about looped models. Thinking Machines also released a research variant, TML-d24-varloop-research (also Apache 2.0), that can run one to five loops, and measured it. Its no-CoT time horizon rose from 16 seconds at one pass to 37 seconds at three, then went flat. On a pure serial lookup task it kept improving out to ten loops, twice what it was trained on. The visible chain of thought got shorter with more loops but stayed legible, and agentic sabotage scores did not rise. Their conclusion is careful: at this scale, no meaningful loss of monitorability.
The economics explain why looping will spread. A non-looping version cost 7.5% less compute to train but needed 140% more inference compute to match the looped model's loss on 512k-token agentic traces. Stanford's Hazy Research and Scaling Intelligence labs measured 0.703 on intelligence per joule on a DGX Spark, against 0.430 for Qwen3.6-35B-A3B.
My read: an open-weights agent coder that runs on a laptop is now ordinary news. What is not ordinary is a lab shipping the safety question together with the model. Thirty-seven seconds of hidden reasoning is harmless today. The curve is what to watch: if looping becomes the cheap way to buy reasoning, more of the thinking moves somewhere no monitor can read, and this release is the first public baseline for measuring that.
← Back to all articles
The reported numbers are strong for its size. DeepSWE v1.1 48.2% against 42.2% for Qwen3.8-27B, SWE-bench Verified 79.2% bash-only, SWE-bench Pro v2 Hard 77.9%, BrowseComp 79.0%, MCP Atlas 82.7%. It does not win everywhere. It trails Qwen3.8-27B on Terminal Bench v2.1 (71.6% vs 73.0%), ties it on the Artificial Analysis index at 20.1, and is weaker at multi-turn chat and many-constraint instructions. Looping adds reasoning, not knowledge, so factual recall tracks its size, and it makes up the gap with search and tools. All of these are vendor-reported.
The bigger story is monitorability. Chain-of-thought monitoring works because a normal transformer can only do so much computation per token before it has to write something down. Looping raises that limit, which is exactly why safety people worry about looped models. Thinking Machines also released a research variant, TML-d24-varloop-research (also Apache 2.0), that can run one to five loops, and measured it. Its no-CoT time horizon rose from 16 seconds at one pass to 37 seconds at three, then went flat. On a pure serial lookup task it kept improving out to ten loops, twice what it was trained on. The visible chain of thought got shorter with more loops but stayed legible, and agentic sabotage scores did not rise. Their conclusion is careful: at this scale, no meaningful loss of monitorability.
The economics explain why looping will spread. A non-looping version cost 7.5% less compute to train but needed 140% more inference compute to match the looped model's loss on 512k-token agentic traces. Stanford's Hazy Research and Scaling Intelligence labs measured 0.703 on intelligence per joule on a DGX Spark, against 0.430 for Qwen3.6-35B-A3B.
My read: an open-weights agent coder that runs on a laptop is now ordinary news. What is not ordinary is a lab shipping the safety question together with the model. Thirty-seven seconds of hidden reasoning is harmless today. The curve is what to watch: if looping becomes the cheap way to buy reasoning, more of the thinking moves somewhere no monitor can read, and this release is the first public baseline for measuring that.
Comments