Thinking Machines’ Tiny Inkling 2 Mini Out-Codes Bigger Open Models

Human Verified

Share to: , and

A laptop on an airplane tray table with a glowing spiral of light rising from the screen

Thinking Machines Lab has released Inkling 2 Mini, a small open-weights model that leads every open model under 100 billion parameters on agentic coding benchmarks. The 26.4-billion-parameter model runs offline on a laptop, using a looped design that replays its middle layers three times per token.

The model scores 48.2% on DeepSWE v1.1 and 46.5% on the Vals Index, the best results among open-weights models of its size. Its total memory footprint is about 20 gigabytes, and the lab says its researchers ran it offline on a MacBook Pro during a flight.

Looping Instead of Growing

Inkling 2 Mini is a 24-layer model, and its two middle blocks of six layers each run three times for every token. The lab says looping lets the model reason longer without writing its reasoning down, which matters for monitoring what the model does by reading its chain of thought.

The design builds on earlier work on looped transformers, including the 2018 Universal Transformer paper. Thinking Machines says looping the middle 12 layers beat looping the last 12 on 27 of 28 tests, and a variable-loop research variant is being released alongside the main model.

The model is a Mixture-of-Experts design with 26.4 billion total parameters, 3.4 billion active per token and 5.4 billion weight activations once looping is counted. Most weights are trained and deployed natively in NVFP4, with the last four layers in MXFP8, keeping the footprint small enough for local devices.

The lab also tested efficiency per watt with Stanford researchers, and Mini scored 0.703 on intelligence per joule, far ahead of Qwen3.6-35B at 0.430 and Gemma-4-26B at 0.159. For a model meant to run on laptops, doing more with less power is the whole point.

Our Read

The interesting claim is not the benchmark score but the method. Looping buys reasoning time without more parameters, and the lab’s tests show the chain of thought stays legible as loops increase. If small models can reason by replaying layers, the race shifts from size to layer use.

The limits are stated honestly, which is rare for a launch post. The model trails larger models on reasoning-heavy tests, and it is not the strongest small model for multi-turn chat or instructions with many constraints. That candour makes the coding lead more believable, not less.

Why It Matters

A laptop-sized model that leads open models on agentic coding changes who can run coding agents. Developers no longer need a data centre to test a capable agent, and researchers can study a frontier-style model locally. The open weights, released under Apache 2.0, make that access real rather than promised.

Background

Inkling 2 Mini follows Inkling 2 Small, the lab’s larger closed model from earlier in November, and its strengths mirror that model’s long-horizon agentic work. The difference is access: Mini is open, small and local, while Small stays behind the lab’s own service.

The launch lands as open models close the gap with closed ones. Alibaba’s Qwen4-Max arrived this month as a round-the-clock AI teammate, and Kimi’s K3.1 Pro put three trillion parameters on a desk. Mini’s answer is different: less size, more thought per token.

Source: Thinking Machines’ Inkling 2 Mini post and The Decoder’s report

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.