Innovation & Tech

Thinking Machines Lab Ships Inkling 2 Mini, a 26.4B Open-Weights Model for Laptops

Thinking Machines Lab's Inkling 2 Mini (26.4B params, ~20GB) runs offline on laptops and beats Qwen3.8-27B on coding benchmarks — but its looped reasoning complicates chain-of-thought monitoring.

By Daniel Okafor

4 min read

Thinking Machines Lab Releases Inkling 2 Mini for Local Devices
Thinking Machines Lab Releases Inkling 2 Mini for Local Devicesjuno1412 / pixabay

What's News

  • Inkling 2 Mini shipped November 20, 2026, with 26.4B total parameters and a ~20GB footprint for consumer laptops.
  • The looped model needed 140% less inference compute than an unpruned non-looping version to match loss on 512,000-token agentic traces.
  • It scores 77.9% on SWE-bench Pro v2 Hard versus 73.8% for Qwen3.8-27B.
  • 24 stored layers produce 48 layer applications per token via three passes through shared middle blocks.
  • Variable-loop tests showed no-chain-of-thought time horizon rising from 16 to 37 seconds between one and three passes.

By Daniel Okafor

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026 — an open-weights model with 26.4 billion parameters that runs on consumer laptops with a roughly 20GB memory footprint. The model repeats passes through its middle layers to reason harder without growing larger.

The full weights are on Hugging Face, alongside a companion research model, TML-d24-varloop-research. Tinker provides inference and fine-tuning, with chat available in Tinker Playground. Local options include Ollama on Mac systems with at least 32GB of RAM, plus GGUF versions through Unsloth and llama.cpp.

In a company demonstration, Inkling 2 Mini ran offline on a MacBook Pro in airplane mode during a flight and built a playable flight simulator in a few minutes. A second demo used one offline MacBook to power all 20 residents of Thinkyville, a simulated pixel town, showing how multiple persistent agents can run locally.

How does the looped architecture work?

Inkling 2 Mini is a Mixture-of-Experts transformer: only a subset of its specialized components fires for each token. Its 24 layers split into four blocks of six. The first and last blocks run once. The two middle blocks share weights and run three times, producing 48 layer applications per token from 24 stored layers. The model has 3.4 billion active parameters and 5.4 billion weight activations per token once looping counts.

Attention residuals let each layer learn its own mixture of earlier block outputs rather than summing everything into a single running total. The second and third passes also receive a gated message from the preceding pass. That gate starts closed and learns to open during training, which lets looping be introduced without destabilizing the model.

Training began with 384 experts, eight selected per token. Two-thirds were pruned during midtraining, leaving the 128 most-used experts per layer. The model then trained on 1 trillion tokens distilled from a larger internal Inkling model to recover performance. Twenty layers deploy in 4-bit NVFP4; the final four run in 8-bit MXFP8. The company says matching training and deployment precision avoids conversion loss from post-training quantization.

What is the compute trade-off?

In the company's training comparison, an unpruned non-looping model cost 7.5% less compute to train but required 140% more inference compute to match the looped version's loss on 512,000-token agentic traces.

The mechanism matters for buyers: looping improves reasoning but does not let weights store more facts, because knowledge tracks total parameter count, the company says. Inkling 2 Mini's AA-Omniscience result matches other models of its size. Its leads among similar-sized open-weights models on BrowseComp, DeepSearchQA and MCP Atlas come from efficient use of search and tools.

Where does the model win and lose?

Thinking Machines Lab reports these coding benchmarks:

  • 48.2% on DeepSWE v1.1 (Qwen3.8-27B: 42.2%)
  • 77.9% on the 51-task SWE-bench Pro v2 Hard (Qwen3.8-27B: 73.8%)
  • 46.5% on Vals Index v2.2

The company notes evaluation setups vary by model. It also flags weaknesses plainly. Inkling 2 Mini is not the strongest small model for multi-turn chat or constraint-heavy instructions, scoring 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1. It trails larger models on reasoning-intensive benchmarks including CritPt v2 and HLE-Diamond.

Why does looping complicate monitoring?

Chain-of-thought monitoring means reading a model's written reasoning to oversee its behavior. A conventional transformer must externalize reasoning that needs extra sequential steps. A looped model can run more of those steps internally before writing anything. Reasoning that never enters the text cannot be inspected by that monitor.

Because Inkling 2 Mini supports only a fixed pass count, the lab trained TML-d24-varloop-research with one to five passes, mostly three, and ran reinforcement learning at three. On a Think Fast subset, the no-chain-of-thought time horizon — the human completion time of tasks the model finishes at 50% reliability without written reasoning — rose from 16 seconds at one pass to 37 seconds at three, then roughly flattened. N-Hop Lookup accuracy kept improving through ten loops, even though training never exceeded five passes.

On GPQA Diamond questions answered correctly at every tested depth, average reasoning traces fell from 3,703 tokens at one pass to 1,661 at five. Mean sentence length barely moved, from 8.9 to 8.7 words, and dictionary-word share held near 94.5%. The shorter traces were more direct and involved less backtracking.

The agentic tests used a helpful-only fine-tune of both models in LinuxArena. With Qwen3.6-35B-A3B as a weaker monitor, the company found no clear trend in stealth rate as loop count rose; restricting the monitor to tool calls and outputs left a significant number of attacks undetected.

The company calls these monitorability results preliminary. Reinforcement learning ran only at three passes, so the experiments do not establish what a model trained at higher pass counts would learn — and the company cautions such a model might compress its reasoning more aggressively. The open question for oversight researchers is whether the legibility seen at this scale survives larger models and training regimes that optimize deeper loops. The fixed and variable open-weights releases give them a direct way to test it.

Source: unite.ai

Share this article:

More from Daniel Okafor

Daniel Okafor

Show full bio

Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.

2,538 articles

Related articles

« Previous article