HomeNews

Inkling 2 Mini: 26.4B Looped Transformer Runs on MacBook

Priyanka Singh
Priyanka Singh
•
Nov 21, 2026 3:46 AM
•
0
 min read
Select Emergent as your Preferred news source
Inkling 2 Mini: 26.4B Looped Transformer Runs on MacBook

💡 TL;DR

  • Thinking Machines released Inkling 2 Mini on November 20, 2026, a 26.4B-parameter open-weights transformer with looped middle layers executing three times per token.
  • The model achieves 48.2% on DeepSWE v1.1 and 77.9% on SWE-bench Pro v2 Hard, leading open-weights models under 100B parameters on agentic coding tasks.
  • With a 20GB memory footprint, Inkling 2 Mini runs on a 32GB MacBook Pro and is available via vLLM, Ollama, and SGLang for local deployment.

Mira Murati's Thinking Machines Lab officially launched on November 20, 2026, Inkling 2 Mini, a 26.4-billion-parameter open-weights transformer designed to run on consumer laptops. The model introduces a looped architecture in which the middle two blocks of six layers each execute three times per token, delivering 5.4 billion effective weight activations while maintaining just 3.4 billion active parameters. This design leverages the hypothesis that middle layers carry abstract reasoning, allowing the model to perform more computation internally before generating output tokens.

Inkling 2 Mini leads every open-weights model with fewer than 100 billion total parameters on key agentic coding benchmarks, including 48.2% on DeepSWE v1.1, 77.9% on SWE-bench Pro v2 Hard, and 46.5% on Vals Index v2.2. The model's 20GB memory footprint enables deployment on devices like a 32GB MacBook Pro, where researchers demonstrated it building a playable flight simulator offline during a commercial flight.

Looped Architecture and Reasoning Gains

The core innovation in Inkling 2 Mini is its looped transformer design. The 24-layer model divides into four blocks of six layers: the first and last blocks execute once, while the two middle blocks run three times per token. This approach builds on prior work in universal transformers and reflects the intuition that middle layers handle more abstract reasoning, while outer layers translate between embeddings and internal representations.

Thinking Machines tested this hypothesis by comparing loop positions across 28 evaluations. Looping the middle 12 layers produced larger accuracy gains than looping the last 12 on 27 of those tests. The model also employs attention residuals, allowing each layer to read outputs from earlier blocks and learn its own mix of them, enabling richer information flow between loops.

A variable-loop research variant released alongside Inkling 2 Mini showed that performance on serial reasoning benchmarks continued improving out to ten loops, twice the number used during training. The no-chain-of-thought time horizon extended from 16 seconds at one pass to 37 seconds with three passes, while generated reasoning remained legible throughout. Critically, tests of agentic sabotage showed no upward trend as loops increased, suggesting the architecture does not amplify alignment risks.

Benchmark Performance and Efficiency

Inkling 2 Mini excels on long-horizon agentic tasks. It leads comparable open-weights models on agentic coding benchmarks and approaches the score of the much larger Inkling 2 Small on PMPP-Hard, a kernel-optimization benchmark. On knowledge-intensive evaluations like AA-Omniscience, the model scores in line with other 26B-parameter models, as looping improves reasoning but does not expand factual storage capacity.

The model compensates for this knowledge gap through efficient search and tool use, leading open-weights peers on BrowseComp, DeepSearchQA, and MCP Atlas. Stanford researchers measuring intelligence per joule found Inkling 2 Mini scored 0.703 on NVIDIA DGX Spark, significantly ahead of Qwen3.6-35B-A3B at 0.430, Gemma-4-26B-A4B at 0.159, and Qwen3.5-27B at 0.036. This efficiency makes the model viable for running multiple persistent agents on local hardware.

Weaknesses include multi-turn chat and constraint-heavy instructions, where Inkling 2 Mini scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, trailing Qwen3.8-27B on both. It also lags behind larger models on reasoning-intensive benchmarks such as CritPt v2 and HLE-Diamond.

Training Pipeline and Quantization

Inkling 2 Mini derives from Thinking Machines' mainline predictive scaling ladder. The team pruned two-thirds of the experts to reduce parameter count and introduced looping to boost reasoning. The model began training with 384 experts at 48 sparsity, selecting 8 experts per layer. At a midtraining checkpoint, the 128 most-used experts in each layer were retained, cutting total parameters and enabling faster inference on local devices.

After pruning, the model trained on 1 trillion tokens distilled from a larger internal Inkling model, recovering pre-pruning performance by teaching remaining experts to absorb the work of deleted ones. Supervised fine-tuning followed with a data mix similar to Inkling 2 Small, then reinforcement learning totaling 2.2 trillion tokens over 2,000 steps, including on-policy distillation for alignment.

Most weights were trained natively in 4-bit NVFP4 precision, with the final four layers in 8-bit MXFP8. The model deploys in the same numeric precision it trained on, eliminating conversion loss from post-training quantization. This approach relies on a dedicated kernel hardware-accelerated on NVIDIA's GB300 platform.

Chain-of-Thought Monitoring and Oversight

A key research question for looped transformers is how they affect chain-of-thought monitoring, a primary tool for understanding and overseeing AI systems. Standard non-looped architectures limit sequential computation steps before a token must be output, forcing reasoning that requires more steps to appear in generated text. Looping raises this limit, enabling models to perform more internal computation before verbalizing their thought processes.

Thinking Machines observed that the variable-loop model's chain of thought grew shorter with more loops but remained legible. Performance on long-context benchmarks like GraphWalks BFS and MRCR v2 improved with additional passes, while knowledge-intensive benchmarks like AA-Omniscience and SimpleQA Verified showed no gains, confirming that looping enhances reasoning without expanding factual storage.

The researchers emphasize that looping does not inherently reduce monitorability. While the model performs more computation per token, its generated reasoning remains interpretable, and no increase in deceptive or misaligned behavior was detected as loop depth increased.

Release Date and Availability

Inkling 2 Mini was officially released on November 20, 2026, with open weights available on Tinker for inference and customization. The model supports deployment via vLLM, Ollama, and SGLang, and runs on consumer devices including 32GB MacBook Pros and desktop AI computers.

What This Means

Inkling 2 Mini demonstrates that looped architectures can deliver strong reasoning performance in a compact, locally deployable package. By running middle layers multiple times, the model achieves gains in agentic coding and long-horizon tasks without requiring frontier-scale compute. The open-weights release and efficient inference profile make it accessible to researchers and developers building on-device AI agents, while the variable-loop research variant provides a testbed for studying the trade-offs between internal computation and external reasoning traces.

About the writer

Priyanka is a Founding Engineer at Emergent, leading the development and evaluation of AI agent systems with a focus on reliability, performance, and scalable user experiences.

Start Building
on Emergent today
Try Emergent