Inkling 2 Mini: a Looped 26B Model That Runs on a Laptop

Thinking Machines' Inkling 2 Mini is an open-weight 26B looped transformer that matches bigger rivals on agentic coding and runs on a MacBook.

@sharedot

Inkling 2 Mini: a Looped 26B Model That Runs on a Laptop

By @sharedot · · 6 pages

  • AI Frontier
  • Open Weights
  • Looped Transformer
  • Agentic Coding

Thinking Machines' Inkling 2 Mini is an open-weight 26B looped transformer that matches bigger rivals on agentic coding and runs on a MacBook.

What Thinking Machines shipped on November 20

Thinking Machines released Inkling 2 Mini on November 20, 2026, an open-weight model with a 1M-token context window and 1M-token output support, published under Apache 2.0 on Hugging Face and Ollama.

The architectural twist is a *looped transformer*: rather than simply scaling parameters, the model runs its middle layers multiple times per token.

Vals AI describes a 26B-parameter model with 5.4B active per token, where twelve of its 24 layers run three times on each token — spending more computation per token instead of more tokens.

Futurum Group, drawing on the company's release, reports a Mixture-of-Experts design with 26.4B total parameters and 3.4B active, fitting a ~20 GB memory footprint small enough for consumer laptops.

The two outlets also differ on how it was demonstrated: Vals AI notes it is small enough to run on a Mac laptop or DGX Spark, while Futurum Group reports researchers ran it offline on a MacBook Pro during a flight, where it built a playable flight simulator in minutes.

Alongside the Mini, the company also shipped a companion variable-loop research model to study what looping does to safety properties.

Why the benchmark numbers are the upset

The surprise is how a ~20 GB model competes with systems far larger.

Futurum Group reports the model scores 48.2% on DeepSWE v1.1 and 79.2% on SWE-bench Verified, leading every open-weights model under 100B total parameters on agentic coding, plus 79.0% on BrowseComp, 85.3% on DeepSearchQA and 82.7% on MCP Atlas on agentic search and tool use.

On Vals' own index the picture is mixed — 46.46% on the Vals Index (#38 of 43), up 17.5 points from the original Inkling, with Vibe Code Bench jumping from 19.21% to 77.92%.

Fans of the efficiency-frontier race will note Mini reaches 96% of its bigger sibling Inkling 2 Small's Vibe Code Bench score and 97% of its SAGE score, per Vals AI — a near-match at a fraction of the size.

Safety and monitorability, with published data

Futurum Group highlights that Thinking Machines released a companion variable-loop research model specifically to study what looping does to safety properties — and the findings are unusual enough to matter.

According to Futurum Group, the model's no-CoT time horizon improved from 16 seconds at one pass to 37 seconds with three passes, and performance on a serial reasoning benchmark kept improving out to ten loops, twice the training depth.

Critically, the chain of thought grew shorter with more loops but stayed legible throughout, and the model showed no increase on tests of agentic sabotage as loops were added.

In a market where Futurum's survey reports 55.4% of enterprise AI decision makers cite agent reliability and hallucination management as a top adoption challenge, monitorability backed by published measurements is a genuine differentiator rather than a marketing claim — a detail practitioners tracking agent reliability will want to read in full.

The stakes: open weights, Tinker, and conversion

The go-to-market may matter as much as the architecture. Futurum Group frames the dual release — open weights on Hugging Face, managed inference and customization on the Tinker platform — as the same playbook hyperscalers used: open weights build mindshare, managed infrastructure captures margin.

The backdrop is favorable but complicated: Futurum reports open-weight models now run in production at 42% of enterprises (up from 31% in July), yet self-hosting costs stall 50% of pilots and only 10% of in-house open-weight builds report sustained ROI at scale.

For builders, the practical hook is a model that clears serious agentic coding bars while running offline on commodity hardware — relevant to the 41.2% of buyers Futurum says rank on-device and edge SLM support among their top vendor criteria.

Futurum's watch list: whether Hugging Face downloads convert into Tinker contracts over Q1 2027, and how Qwen, Gemma and Muse Glimmer respond on DeepSWE and SWE-bench Pro v2 Hard.

Sources

  1. vals.ai › Inkling 2 Mini Benchmarks, Cost and Capabilities
  2. futurumgroup.com › Thinking Machines' Inkling 2 Mini Redraws the Open-Weight Efficiency Frontier

More on AI Frontier