Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, released a new open-weights model on November 20, 2026: Inkling 2 Mini, a compact model built around an unusual architecture that the company says lets it punch well above its parameter count. The model is small enough to run locally on a consumer laptop, with a total memory footprint of roughly 20GB.

The headline numbers are striking for the size class. According to the company's announcement, Inkling 2 Mini leads every open-weights model with less than 100 billion total parameters on DeepSWE v1.1, resolving 48.2 percent of tasks, and on Vals Index v2.2, scoring 46.5 percent. Both benchmarks measure agentic coding ability, one of the most commercially contested capability frontiers in the industry right now. For more context on this story, see our ongoing artificial intelligence updates.

What Makes a Looped Transformer Different

Inkling 2 Mini is a looped transformer. Instead of passing each token through every layer exactly once, the model runs its middle layers three times per token. In effect, the model can apply more sequential computation to a problem without generating more output tokens.

The company's framing is that looping allows the model to reason longer without writing its reasoning down. Traditional chain-of-thought reasoning spends tokens on visible intermediate steps; looping spends compute instead. That has real consequences for speed on tasks where reasoning quality matters more than token count, and agentic coding is exactly that kind of task.

Under the hood, Inkling 2 Mini is a Mixture-of-Experts model with 26.4 billion total parameters and just 3.4 billion active per token, with 5.4 billion weight activations per token once looping is accounted for. Most of the weights were trained and are deployed natively in NVFP4, the low-precision format Nvidia has been pushing across its data center lineup, with the final four layers kept in MXFP8. The combination of aggressive quantization and a small active parameter count is what squeezes the model into a roughly 20GB footprint that fits on laptops and desktop AI machines.

The lab demonstrated the model running offline, in airplane mode, on a MacBook Pro with an M5 Pro chip and 48GB of memory, where it built a playable flight simulator in a few minutes.

The Monitorability Trade-Off

The release is accompanied by research that may matter more than the benchmarks. If a model reasons internally through loops rather than expressing its chain of thought in tokens, then the visible reasoning text stops being a faithful window into what the model is actually doing. That has implications for AI safety research, where monitoring chain-of-thought has become one of the few practical oversight tools available.

Alongside Inkling 2 Mini, Thinking Machines released a research-specific variable-loop variant to measure the effect. The results cut in an encouraging direction for capability, a complicating one for oversight. The variable-loop model's time horizon on tasks completed without any chain of thought improved from 16 seconds at a single pass to 37 seconds at three loops, and performance on a serial reasoning benchmark kept improving out to ten loops — twice as many as the model was trained with.

Meanwhile, the visible chain of thought grew shorter as loops were added, but the company reports it remained legible throughout, and the model did not trend higher on tests of agentic sabotage as looping depth increased. Thinking Machines describes these as some of the first looped-architecture models available to researchers, positioning the release partly as an invitation for the safety community to study monitorability on real weights rather than simulated setups.

A Competitive Play in the Sub-100B Class

The benchmarks place Inkling 2 Mini against a crowded sub-100B field that includes Gemma 4 models in the 26B–31B range, Qwen3.8-27B, Muse Glimmer, and a 65B preview model called Ripple. Thinking Machines claims its model also runs faster than similarly performing competitors, citing theoretical token throughput at batch size one.

The sub-100B segment has become strategically important for two reasons. First, on-device and private deployment is growing fast as enterprises balk at sending sensitive data to hosted APIs. Second, per-token economics for local inference increasingly favor small, aggressively quantized models over distilled copies of large ones. A 26.4B MoE that runs natively in NVFP4 with looping compute is a direct answer to that market.

The model is available on Hugging Face with a published model card, and Thinking Machines is promoting access through its Tinker fine-tuning platform. Between this release and the lab's earlier Inkling family models, the startup is building a distinct technical identity around recurrence-style compute scaling at a time when most frontier labs remain focused on scaling parameters and data.

Whether looping becomes a mainstream architecture or remains a niche, the release marks a notable moment: one of the most closely watched AI startups is betting that smarter computation per token — not more tokens, and not more parameters — is where the next gains live.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →