Thinking Machines Releases Inkling 2 Mini. The Card Says 26.4B Parameters and a 32GB Floor.
Thinking Machines Lab released Inkling 2 Mini on November 20, 2026. The model card lists 26.4 billion total parameters, 3.4 billion active, and 5.4 billion effective active parameters per token once the loop is counted. The checkpoint is about 20GB. The company says it runs on a laptop. The card's floor is 32GB of memory, not a smaller machine.
The numbers
- 26.4 billion total parameters, 3.4 billion active. Each extra pass through the looped layers adds about 1.0 billion effective active parameters, and the model runs three passes, so the card counts 5.4 billion effective active parameters per token. Model card.
- 24 layers. The middle 12 run three times on shared weights, which the card calls 48 layer applications per token. Each token is routed to 8 of 128 experts. Model card.
- About 20GB, mixed NVFP4 and MXFP8. The news post says 20 layers are deployed in NVFP4 and the last four in MXFP8. The public repo thinkingmachines/Inkling-2-Mini is tagged apache-2.0, and its stored size is 20,191,385,756 bytes.
- Context window up to 1 million tokens. Text and image in, text out. Model card.
- License Apache 2.0. Release date on the card: November 20, 2026. Provider: Thinking Machines Lab, Inc.
What "runs on a laptop" means in the card
The November 20 news post says Inkling 2 Mini is small enough to run on a laptop, and that researchers ran it offline on a MacBook Pro and built a flight simulator. The hardware list is more specific. The card says the supported machines are one NVIDIA DGX Spark, Apple silicon with 32GB or more (MacBooks, Mac minis, and Studios), and other GPUs with 32GB or more of VRAM recommended. The demo caption names a MacBook Pro with an M5 Pro and 48GB, in airplane mode.
On Apple silicon the card names Ollama and Exo Labs. On NVIDIA it names SGLang, vLLM, Unsloth, NVIDIA Dynamo, and Tokenspeed. It also says quantized GGUF versions are available via Unsloth and llama.cpp. The weights are on Hugging Face, and the post says the model is on Tinker for inference and customization.
The loop, and what the company measured
Inkling 2 Mini is a looped transformer. The news post says the two middle blocks of six layers run three times, and the first and last blocks run once. Thinking Machines says looping the middle 12 layers beat looping the last 12 on 27 of 28 evaluations. A non-looping comparison model cost 7.5 percent less compute to train and needed 140 percent more inference compute to match the looped model's loss on 512,000-token agentic traces. Those figures are the lab's.
The post also releases a research model, TML-d24-varloop-research, because the shipping Mini uses a fixed number of passes. That model is public on Hugging Face under the same apache-2.0 tag. On it, Thinking Machines says the no-chain-of-thought time horizon moved from 16 seconds at one pass to 37 seconds at three. The chain of thought got shorter as loops were added and stayed legible, and the model did not trend higher on the lab's tests of agentic sabotage. This is an internal monitorability note, not an outside audit.
Scores the lab prints, and where it says the model loses
Thinking Machines says Inkling 2 Mini leads every open-weights model under 100 billion total parameters on DeepSWE v1.1 at 48.2 percent and on Vals Index v2.2 at 46.5 percent. On its table, AA Index v5.0 is 20.1, tied with Qwen3.8-27B. SWE-bench Verified, bash only, is 79.2 percent. Terminal Bench v2.1 is 71.6 percent, behind Qwen3.8-27B at 73.0 percent. The same table marks those as the lab's reported scores. The chart note says evaluation setups differ by model.
The limitations section is direct. Inkling 2 Mini scores 50.0 percent on MultiChallenge and 33.4 percent on ComplexConstraints v1.1, behind Qwen3.8-27B on both. The post says looping does not store more facts, and AA-Omniscience v1.1 is -8.6, behind Nemotron 3.5 Super VL at -2.9 and Qwen3.8-27B at -6.4.
What is not disclosed?
The card does not print a dollar price for Tinker or for third-party APIs. It does not say how many downloads the weights had on release day. The Hugging Face API later showed a stored size and a parameter count in the safetensors header, 26,388,646,440, which rounds to the card's 26.4 billion. That header does not break out the 3.4 billion active figure. The post does not name the MacBook's tokens per second. A social-media cluster on November 22 is not the source for any number here. The source is the November 20 post and the model card.
Quick answers
- What happened with Thinking Machines Lab?
- Thinking Machines Lab released Inkling 2 Mini on November 20, 2026. The model card lists 26.4 billion total parameters, 3.4 billion active, and 5.4 billion effective active parameters per token once the loop is counted. The checkpoint is about 20GB. The company says it runs on a laptop. The card's floor is 32GB of memory, not a smaller machine.
- What are the key numbers?
- 26.4 billion total parameters, 3.4 billion active. Each extra pass through the looped layers adds about 1.0 billion effective active parameters, and the model runs three passes, so the card counts 5.4 billion effective active parameters per token. Model card.
- What is not disclosed?
- The card does not print a dollar price for Tinker or for third-party APIs. It does not say how many downloads the weights had on release day. The Hugging Face API later showed a stored size and a parameter count in the safetensors header, 26,388,646,440, which rounds to the card's 26.4 billion. That header does not break out the 3.4 billion active figure. The post does not name the MacBook's tokens per second.
This article was ultrathought.