Back to News
ProductAI Understanding briefing

Thinking Machines releases Inkling 2 Mini, a 26B looped transformer that runs locally on laptops

Thinking Machines Lab unveiled Inkling 2 Mini, a 26.4 billion‑parameter open‑weights multimodal model that can run offline on a 32 GB MacBook Pro or a single GPU, promising faster, cheaper agentic coding performance.

4 min readRead the original reporting
Source-provided image accompanying Thinking Machines releases Inkling 2 Mini, a 26B looped transformer that runs locally on laptops
Attributed reportingSource recorded
Publisher
venturebeat.com
Source type
Reporting by a news outlet — not a first-party document.

What we could not confirm independently: This claim is attributed to the named outlet. We did not verify it against a first-party document. (venturebeat.com)

ContextUnderstand this in 60 seconds

Key terms

Transformer
A neural architecture that uses attention to model relationships across sequences in parallel.
Mixture of Experts (MoE)
An architecture with specialized subnetworks where only selected experts run per input.
Memory (Agent Memory)
Stored context an AI agent uses across steps or sessions to improve continuity.

What happened

Thinking Machines Lab launched Inkling 2 Mini, a 26.4 billion‑parameter, 20 GB‑footprint multimodal model that runs locally on consumer‑grade hardware. The model uses a novel “looped ” architecture, repeatedly processing its middle layers to achieve effective depth without increasing storage. Weights and model card were released on Hugging Face under an Apache 2.0 license, and early benchmarks show it outperforming comparable open‑weight models on software‑engineering and tool‑use tasks while costing less compute per task.

Thinking Machines Lab announced Inkling 2 Mini on November 5, 2026. The model contains 26.4 billion total parameters, with a 20 GB memory footprint that fits on a 32 GB MacBook Pro or a single desktop GPU. It is a 24‑layer decoder‑only sparse Mixture‑of‑Experts (MoE) model that routes each token to 8 of 128 experts, activating 3.4 billion parameters per pass. The middle 12 layers are executed three times with shared weights, yielding 48 effective layer applications per token and an effective 5.4 billion active parameters without increasing storage.

The looped architecture is intended to boost sequential reasoning without the token‑inflation of chain‑of‑thought prompting. Internal tests reported a 140 % reduction in compute needed to match loss on long‑token traces compared to a non‑looping baseline. The model was trained with 4‑bit NVFP4 precision for the first 20 layers and 8‑bit MXFP8 for the final four, preserving training‑time numeric formats to avoid post‑training quantization loss.

Inkling 2 Mini’s performance was benchmarked on several open‑weight leaderboards. It leads sub‑100 B models on DeepSWE v1.1 (48.2 %), SWE‑bench Pro v2 Hard (77.9 %), Vals Index v2.2 (46.5 %), and scores 79.2 % on SWE‑bench Verified and 85.3 % on DeepSearchQA. On the Stanford intelligence‑per‑joule test, it achieved 0.703, outpacing Qwen 3.6‑35B‑A3B (0.430) and Gemma‑4‑26B‑A4B (0.159).

The model’s limitations were also disclosed: factual recall remains lower than denser peers, with a -8.6 score on AA‑Omniscience v1.1 and a 63.4 % non‑hallucination rate. Multi‑turn chat and complex‑constraint instruction scores trail Qwen 3.8‑27B, and occasional repetitive loops or clipped phrasing were observed in non‑English outputs.

Weights and the model card were made publicly available on Hugging Face under an Apache 2.0 license, accompanied by a standard Model Acceptable Use Policy. Early community feedback highlighted strong local‑inference performance on tools like llama.cpp and vLLM, with reported decode speeds of 210 tokens/sec on a 96 GB M5 Ultra Mac Studio and successful offline demos such as building a flight‑simulator app in airplane mode.

Source details: venturebeat.com

Why it matters

Inkling 2 Mini demonstrates that high‑agency AI can be deployed offline on laptops, reducing reliance on costly cloud APIs and addressing data‑privacy concerns for enterprises. Its looped design offers a new efficiency pathway, achieving comparable benchmark scores to larger 27‑30 B models while using roughly a third of the compute. The open‑source Apache 2.0 licensing removes legal barriers for commercial integration, potentially accelerating adoption of private, on‑premises AI assistants in software development, simulation, and other high‑value domains. However, the model still lags on factual recall and multi‑turn chat, highlighting trade‑offs between compactness and knowledge breadth.

Local deployment reduces dependence on cloud providers, cutting operational costs and mitigating data‑privacy risks for organizations that need to keep proprietary code or sensitive information on‑premises.

The looped architecture offers a novel efficiency trade‑off, showing that deeper reasoning can be achieved without scaling parameter counts, which could influence future model design across the industry.

Apache 2.0 licensing removes legal friction for commercial use, enabling startups, research labs, and large enterprises to embed the model in proprietary products without copyleft obligations, potentially expanding the ecosystem of private AI assistants.

Benchmark results suggest Inkling 2 Mini can handle agentic coding and tool‑use tasks at a fraction of the compute cost of larger models, making high‑performance AI more accessible to developers with modest hardware.

Safety considerations remain, as the looping mechanism reduces the visibility of internal reasoning steps. The release of a variable‑loop checkpoint and partnerships with interpretability groups indicate an emerging focus on monitoring and oversight for such architectures.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Interactive Concept Check+10 Points
What is AI? Quiz

A route planner searches possible journeys using explicit rules. What does this illustrate about AI?

What to watch next

Future iterations of looped models, adoption rates of Inkling 2 Mini in enterprise workflows, and any emerging safety or interpretability findings from the variable‑loop research checkpoint.

Adoption metrics: how quickly developers and enterprises integrate Inkling 2 Mini into local workflows, especially for code generation and autonomous simulation tasks.

Further research on looped safety: whether the reduced external reasoning traces affect model interpretability, alignment, or susceptibility to hidden failure modes.

Performance evolution: upcoming releases from Thinking Machines or competitors that may refine looped architectures or address the model’s factual recall gaps.

Commercial licensing activity: whether the permissive Apache 2.0 terms lead to notable enterprise deployments or inspire similar licensing choices for other open‑weight models.

Related guides & quizzes

Found this useful?