What Thinking Machines released
Thinking Machines Lab released Inkling 2 Mini on 20 November 2026 with open weights. The checkpoint is on Hugging Face under the Apache 2.0 licence. Inkling 2 Mini is a mixture-of-experts (MoE) model, meaning each token goes through only a small subset of specialist feed-forward blocks. It has 26.4 billion parameters in total, and 3.4 billion of them are active for each token. The lab puts the whole checkpoint at about 20GB of memory. The model card lists two supported setups: a single NVIDIA DGX Spark, or an Apple silicon machine with at least 32GB of memory. The model accepts text and image input and produces text, with a context window of up to 1 million tokens.
Builders in Auckland will care about the combination. The licence is permissive, the lab says the agentic coding scores lead open models under 100B parameters, and the memory footprint suits a high-spec MacBook Pro or Mac mini that plenty of developers already own.
How a looped transformer works
A standard transformer sends each token through its stack of layers once. Inkling 2 Mini stores 24 layers in four blocks of six. The first and last blocks run once per token. The two middle blocks run three times using the same weights, which gives each token 48 layer applications from 24 stored layers. The announcement counts the repeated passes as 5.4 billion weight activations per token. That difference explains why The Decoder reports 5.4B active parameters while the lab's own figure is 3.4B.
A learned gate decides how much of each pass feeds into the next one. At the start of training the gate blocks the previous pass completely, so looping can switch on without destabilising the model. Thinking Machines reports that looping the middle 12 layers beat looping the last 12 on 27 of 28 evaluations. The lab also says a non-looped equivalent needed 140% more inference compute to reach the same loss on 512k-token agentic traces. The Next Web reports the training-side cost of that choice: the non-looped version was only 7.5% cheaper to train.
Two other design choices keep the model small. Training began with 384 experts per layer. During midtraining the lab cut each layer down to its 128 most-used experts, then recovered quality by distilling 1 trillion tokens from a larger internal Inkling model. Most weights use NVIDIA's 4-bit NVFP4 format, and the last four layers use MXFP8. Because the model was trained in those formats, the lab says no quality is lost converting from a higher-precision checkpoint. Post-training used supervised fine-tuning and then 2.2 trillion tokens of reinforcement learning.
Benchmarks and what they cover
The lab's headline results are 48.2% on DeepSWE v1.1, 46.5% on the Vals Index v2.2, 79.2% on SWE-bench Verified and 71.6% on Terminal Bench v2.1. In a direct comparison with Qwen3.8-27B, The Next Web lists 48.2% against 42.2% on DeepSWE and 77.9% against 73.8% on a hard SWE-bench Pro set. Thinking Machines produced all of these numbers itself.
Artificial Analysis provides an outside measurement. Inkling 2 Mini scored 20 on its Intelligence Index, the same as Qwen3.8 27B. The Next Web reports that, in Artificial Analysis's runs, Inkling 2 Mini cost $0.59 and took 9 minutes per task, against $3.53 and 38.2 minutes for Qwen3.8 27B. Those costs depend on the hosting used for the evaluation, so expect different numbers on your own hardware.
Limits the lab admits
Thinking Machines is open about the model's weak spots. The announcement says Inkling 2 Mini "is not the strongest small model for multi-turn chat or for instructions with many constraints". Qwen3.8-27B beats it on MultiChallenge, where Inkling 2 Mini scores 50.0%, and on ComplexConstraints v1.1, where it scores 33.4%. Looping spends more compute on each token but stores no extra knowledge, so factual recall is about what you'd expect from a 26B model.
The model card gives agent builders some practical warnings. During long debugging sessions Inkling 2 Mini can get stuck re-checking its own work, so the card recommends capping output tokens and agent turns. Inkling 2 Mini handles tools less reliably than Inkling 2 Small. It may delete or overwrite files without asking for confirmation and then misreport the result. Output quality drops in languages other than English. Local runtimes have only just added support for looped models, so token speed varies from one tool to another.
Running it locally
At launch the model card lists Ollama and Exo Labs for Apple silicon. For NVIDIA hardware it lists SGLang, vLLM, Unsloth, NVIDIA Dynamo and TokenSpeed. The Ollama library page shows three 20GB tags, one of them an MLX build, and a one-line command:
ollama run inkling-2-miniOn a Linux machine with a suitable NVIDIA GPU, the Hugging Face page shows how to serve the model with vLLM, which exposes an OpenAI-compatible endpoint:
pip install vllm
vllm serve "thinkingmachines/Inkling-2-Mini"For GPUs other than the DGX Spark, the model card recommends at least 32GB of VRAM. If your card has 24GB or less, you'll need one of the GGUF quantisations from Unsloth or llama.cpp, which make the model smaller at some cost to quality. For fine-tuning, the lab points to Tinker, its hosted training API, and to the open reinforcement learning frameworks SkyRL, Miles and NVIDIA NeMo AutoModel. The Tinker playground lets you try the model before you download 20GB over a home connection.
The lab also released a companion research model, TML-d24-varloop-research, which lets you change the number of loops at inference time. Thinking Machines reports that accuracy at each reasoning-token budget rises as loops increase, with diminishing returns after three passes. If you want to measure how loop count trades speed against quality on your own tasks, use the research model.
What NZ builders could do with it
Inkling 2 Mini needs 20GB, has a 1M-token window and comes under Apache 2.0. That makes it a fit for work where data has to stay on your own machines. One example is a coding assistant that works over a private repository. Another is document extraction for a client who won't accept offshore processing. Because the model runs offline, it could also support field work in parts of the country with poor connectivity.
The model card's safety advice sets out the precautions. Run agent workloads in a sandbox and require confirmation before any destructive file or shell action. Put hard limits on turns and output tokens to stop the re-checking loops the card describes. The lab's coding scores justify a trial, but test the model on your own repositories before you replace a hosted model with it. If you get Inkling 2 Mini running, the society's projects page is a good place to post your token speeds and memory use so other members can compare hardware.
Sources
- Introducing Inkling 2 Mini - Thinking Machines Lab
- Inkling 2 Mini model card - Thinking Machines Lab
- thinkingmachines/Inkling-2-Mini - Hugging Face
- inkling-2-mini - Ollama library
- Thinking Machines releases Inkling 2 Mini, a looped AI model for laptops - The Next Web
- New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop - The Decoder
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.