Skip to main content
Live
Main content

Thinking Machines releases Inkling 2 Mini, a 26.4B open-weights model for laptops

The looped mixture-of-experts model activates 3.4B parameters per token, fits in about 20GB of memory and scores 48.2% on DeepSWE v1.1.

Jaeden Schafer
Editor in Chief · · 5 min read
Thinking Machines releases Inkling 2 Mini, a 26.4B open-weights model for laptops

Thinking Machines Lab released Inkling 2 Mini on November 20, an open-weights model built to run on a laptop, with 26.4 billion total parameters and 3.4 billion active per token. The weights are available on Hugging Face, and the model can also be used for inference and fine-tuning on the company's Tinker platform, with a chat interface in Tinker Playground.

The release answers feedback on the lab's previous model. Thinking Machines said some users found Inkling 2 Small too big at 594 billion parameters. Inkling 2 Mini has a memory footprint of about 20GB, helped by native low-precision training: most weights use the 4-bit NVFP4 format, with the final four layers in 8-bit MXFP8.

Inkling 2 Mini is a looped mixture-of-experts transformer. It stores 24 layers in four blocks of six. The first and last blocks run once, while the two middle blocks share weights and run three times, for 48 layer applications per token. Counting the loops, the model performs 5.4 billion weight activations per token.

Key facts

  • 01Inkling 2 Mini has 26.4B total parameters, 3.4B active per token, and a memory footprint of about 20GB.
  • 02It is a looped transformer: 24 stored layers, with the middle 12 run three times for 48 layer applications per token.
  • 03It scores 48.2% on DeepSWE v1.1 versus 42.2% for Qwen3.8-27B, and 77.9% on SWE-bench Pro v2 Hard versus 73.8%.
  • 04Weights are on Hugging Face; the model is also on Thinking Machines' Tinker platform for inference and fine-tuning.
  • 05It follows Inkling 2 Small, a 594B-parameter model some users found too big to run themselves.

Thinking Machines said looping the middle layers produced a 1.51x to 1.91x compute-equivalent gain, compared with 1.24x to 1.36x for looping only the last layers. A non-looping model cost 7.5% less training compute but needed 140% more inference compute to match the looped model on 512,000-token agentic traces. Mira Murati, who leads the lab, summed up the tradeoff on X.

“Looping lets a model think more and write less.”
— Mira Murati, CEO of Thinking Machines Lab

Training started with 384 experts, with 8 selected per token. During midtraining, Thinking Machines pruned two-thirds of the experts, keeping the 128 most-used per layer, then healed the model on 1 trillion tokens distilled from a larger internal Inkling model. Post-training combined supervised fine-tuning, reinforcement learning totaling 2.2 trillion tokens over 2,000 steps, and on-policy distillation.

On the company's benchmarks, Inkling 2 Mini scored 48.2% on DeepSWE v1.1, ahead of Qwen3.8-27B at 42.2%, and 77.9% on SWE-bench Pro v2 Hard against 73.8% for Qwen3.8-27B. It posted 79.2% on SWE-bench Verified in bash-only mode, 46.5% on Vals Index v2.2, 79.0% on BrowseComp with compaction and 82.7% on MCP Atlas. Thinking Machines said the model leads open-weights models under 100 billion total parameters on DeepSWE and Vals Index.

The lab also listed weak spots. Inkling 2 Mini scored 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, trailing Qwen3.8-27B on multi-turn chat and many-constraint instructions, and 71.6% on Terminal Bench v2.1 versus 73.0% for Qwen3.8-27B. Artificial Analysis put the model at 20.1 on its Intelligence Index, tying Qwen3.8-27B, at $0.59 and 9.0 minutes per task versus $3.53 and 38.2 minutes for the Qwen model.

Thinking Machines showed the model running fully offline. A MacBook Pro in airplane mode on a flight built a playable flight simulator in a few minutes, and a single offline MacBook ran all 20 villagers in a simulated pixel town called Thinkyville. Day-one support covers SGLang, vLLM, Ollama, Unsloth and NVIDIA Dynamo, with Ollama on Macs with at least 32GB of RAM.

Related · from this week
Nvidia launches $4,999 DGX Spark 64GB for local AI development
Jaeden Schafer · 4 min read →

Thinking Machines also measured how looping affects the visible reasoning a model writes out. On GPQA, average reasoning traces fell from 3,703 tokens at one pass to 1,661 tokens at five passes, and the company called its monitorability findings preliminary.

“At this scale, its thought process remains monitorable; this may not stay the case.”
— Mira Murati, CEO of Thinking Machines Lab

Alongside Inkling 2 Mini, the lab released a research model, TML-d24-varloop-research, trained with one to five passes. Its time horizon without chain-of-thought rose from 16 seconds at one pass to 37 seconds at three before roughly flattening.

Thinking Machines said it is working with the National Deep Inference Fabric to bring Inkling 2 Mini to interpretability researchers, and that it gave Goodfire early access to study how looped models work.

ShareXLinkedInEmail
AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

  • ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat
  • Generate images & video with Sora, Veo, Ideogram
  • Compare any two models side by side
  • From $8.99/mo · 80+ models, all included
Try AI Boxaibox.ai
Trusted by 3,000+ teams
Got a tip?

Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.

Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.

AI Box Daily briefingFree · Daily · No fluff

Stay ahead of everyone in AI.

The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.

Loved by 10,000+ AI professionals
Free forever. Unsubscribe with one click.

The briefing read inside teams at

Keep reading

More from Models

Nvidia logo
Models

Nvidia launches $4,999 DGX Spark 64GB for local AI development

The new SKU runs 100B-parameter models on device; two units cluster to 128GB and 200B via Nvidia Sync Cluster Assistant.

Jaeden Schafer4 min read
OpenAI logo
Models

OpenAI's Astra and Anthropic's Claude Opus 5 crack unsolved Enigma messages

Two cryptanalysts used frontier LLMs to break wartime codes that had resisted human researchers for two decades.

Jaeden Schafer4 min read
Anthropic logo
Models

Anthropic says Claude autonomously discovered a Crispr-like enzyme system

Nearly 950 Claude agents worked 21 hours across 210 million tokens to flag a previously uncharacterized enzyme found in bacteriophages.

Jaeden Schafer4 min read